1 00:00:00,240 --> 00:00:02,360 Hi, and welcome to the Neil Ashton Podcast. 2 00:00:03,000 --> 00:00:05,880 In each episode, we explained some of the fascinating ways 3 00:00:05,880 --> 00:00:09,080 that science and engineering are changing the world around us. 4 00:00:09,760 --> 00:00:12,720 We talked to leading engineers from elite level sports like 5 00:00:12,800 --> 00:00:16,840 cycling and Formula One to some of the world's top academics to 6 00:00:16,840 --> 00:00:20,960 understand how fluid dynamics, machine learning, supercomputing 7 00:00:21,360 --> 00:00:22,960 are bringing in a new era of discovery. 8 00:00:23,920 --> 00:00:27,040 We also hear some of their life stories, their career advice, 9 00:00:27,640 --> 00:00:30,040 the lessons they've learned on the way that I hope will be 10 00:00:30,040 --> 00:00:33,800 helpful to you too. So sit back and enjoy this 11 00:00:33,800 --> 00:00:40,840 episode. Hi, and welcome back to the Neil 12 00:00:40,840 --> 00:00:43,160 Ashton Podcast. So today I'm delighted to be 13 00:00:43,160 --> 00:00:45,240 joined by Professor Paola Cinnella. 14 00:00:45,440 --> 00:00:48,240 She's a professor of fluid mechanics at Sorbonne University 15 00:00:48,240 --> 00:00:52,310 in Paris, and she's also the new director of the Sorbonne 16 00:00:52,310 --> 00:00:55,680 Cluster for Artificial Intelligence, also in Paris. 17 00:00:56,280 --> 00:00:59,196 And what I've found particularly interesting about her career 18 00:00:59,196 --> 00:01:02,102 is that she's not come through AI, through the usual, you 19 00:01:02,102 --> 00:01:05,000 know, computer science route. She is very much mechanical 20 00:01:05,000 --> 00:01:07,760 engineering and fluid dynamics specialist. 21 00:01:08,080 --> 00:01:12,440 We talk about where she started in Italy, the transition to 22 00:01:12,440 --> 00:01:17,080 France, moving back, the idea of, you know, academia having to 23 00:01:17,080 --> 00:01:21,840 move between different countries to, to secure the tenure 24 00:01:21,840 --> 00:01:25,200 positions, how how challenging it, it, it can be, especially 25 00:01:25,200 --> 00:01:27,360 for, you know, for family and, and personal life. 26 00:01:28,200 --> 00:01:30,560 But she really, well, I liked it. 27 00:01:30,560 --> 00:01:34,000 Her passion for, for fluids, you know, she talked about working 28 00:01:34,000 --> 00:01:38,520 on these exotic dense gases and real gas effects and, and how 29 00:01:39,080 --> 00:01:43,160 some of those LED her towards the uncertainty quantification 30 00:01:43,160 --> 00:01:46,880 and Bayesian methods. And working with statisticians 31 00:01:47,040 --> 00:01:50,817 helped her then realize that she could do the same around 32 00:01:50,817 --> 00:01:54,310 machine learning and really wanting to collaborate with 33 00:01:54,310 --> 00:01:56,480 different disciplines and, and different groups. 34 00:01:56,880 --> 00:02:00,040 And, and in some ways, she's represents a very important 35 00:02:00,040 --> 00:02:03,720 bridge between what we might call, you know, classical CFD, 36 00:02:03,720 --> 00:02:06,680 you know, numerical methods, turbulence modelling and this 37 00:02:06,680 --> 00:02:09,560 new world of, of, of AI for science. 38 00:02:09,560 --> 00:02:13,720 And, and she's played actually a, a major role in the wider CFD 39 00:02:13,720 --> 00:02:15,200 community. You know, she's an editorial and 40 00:02:15,200 --> 00:02:18,520 chief of Computers & Fluids associate editor for 41 00:02:18,520 --> 00:02:20,200 International Journal of Heat and Fluid Flow. 42 00:02:20,600 --> 00:02:23,840 And she also helps to coordinate the ERCOFTAC special interest 43 00:02:23,840 --> 00:02:28,784 group on Machine Learning in Fluid Dynamics, which has led to 44 00:02:28,784 --> 00:02:30,640 a really successful conference series. 45 00:02:30,640 --> 00:02:34,800 This ML fluids that I have been fortunate to to have helped her 46 00:02:35,240 --> 00:02:40,440 a little bit as well. And, but really we, she is an 47 00:02:40,440 --> 00:02:43,400 educator and she is an academic and, and we had a really 48 00:02:43,400 --> 00:02:48,440 interesting discussion as well on what it means in this new era 49 00:02:48,440 --> 00:02:51,594 of AI, what, what's the valuable thing that we should be 50 00:02:51,594 --> 00:02:54,343 teaching new students at undergraduate and postgraduate 51 00:02:54,343 --> 00:02:57,190 level? And really what is the potential 52 00:02:57,190 --> 00:02:59,860 for methods? And one of the things we talk 53 00:02:59,868 --> 00:03:02,327 about is, you know, the differences between surrogate 54 00:03:02,335 --> 00:03:05,247 modelling and turbulence modelling and how could they be 55 00:03:05,247 --> 00:03:07,554 interconnected? When is the value of those? 56 00:03:07,562 --> 00:03:11,415 And and we finish off talking a little bit on the AI for science 57 00:03:11,415 --> 00:03:14,001 and how there is this opportunity for cross 58 00:03:14,009 --> 00:03:18,135 collaboration across discipline. So she is somebody that I am 59 00:03:18,143 --> 00:03:20,760 inspired by and always enjoy working with. 60 00:03:20,768 --> 00:03:25,063 And I hope you enjoy learning more about her and her work in 61 00:03:25,071 --> 00:03:27,641 this conversation. So sit back and enjoy this 62 00:03:27,649 --> 00:03:29,790 conversation with Paola. So yeah, thanks. 63 00:03:29,798 --> 00:03:32,400 Thanks for coming on this, really appreciate it. 64 00:03:32,408 --> 00:03:35,280 I've loved working with you on various projects over the years, 65 00:03:35,280 --> 00:03:38,520 but this is a good opportunity to hear more about, you know, 66 00:03:38,528 --> 00:03:43,750 you and, and what you do. And I guess maybe it's a 67 00:03:43,758 --> 00:03:46,988 starting question. You know, you started in 68 00:03:46,996 --> 00:03:49,360 mechanical engineering in, in classical fluid mechanics, I 69 00:03:49,368 --> 00:03:52,605 guess, and then you move through now into, you know, machine 70 00:03:52,613 --> 00:03:56,208 learning and, and AI. But how did all that happen? 71 00:03:56,216 --> 00:03:58,980 You know, rewind where? Where did this love for fluid 72 00:03:58,988 --> 00:04:02,584 mechanics start? Well, when I started, so I had 73 00:04:02,592 --> 00:04:06,114 to choose the university. I wanted to do fundamental 74 00:04:06,122 --> 00:04:08,850 physics. I, I wanted to go to, to 75 00:04:08,858 --> 00:04:11,947 fundamental physics, astrophysics or something or to 76 00:04:11,947 --> 00:04:15,455 mathematics. And well, my mother said, no, 77 00:04:15,463 --> 00:04:19,954 this is not a good job because the only opportunity for you is 78 00:04:19,954 --> 00:04:22,326 to become a researcher that's not well paid. 79 00:04:22,334 --> 00:04:25,157 So be an engineer. So good for engineering. 80 00:04:25,165 --> 00:04:28,510 And then I started to do mechanical engineering actually. 81 00:04:28,510 --> 00:04:31,720 And after one year, I discovered 82 00:04:31,720 --> 00:04:35,760 that they didn't like, and I started to look for things which 83 00:04:35,760 --> 00:04:37,960 were closer to physics and to mathematics. 84 00:04:38,400 --> 00:04:44,080 And well, a friend of mine said, you know, I'm doing fluid 85 00:04:44,080 --> 00:04:45,680 mechanics. There's quite a lot of 86 00:04:45,680 --> 00:04:47,720 mathematics in that, and also some physics. 87 00:04:47,720 --> 00:04:51,600 So, right. And that's how I went to fluid 88 00:04:51,600 --> 00:04:53,980 mechanics and I didn't like it. 89 00:04:53,988 --> 00:04:57,720 Actually, my my professor in Italy is Michele Napolitano. 90 00:04:57,728 --> 00:05:02,523 He was, he worked with NASA. He was, he had a PhD with 91 00:05:02,523 --> 00:05:05,472 Stanley Rubin, who was actually the one of the first 92 00:05:05,472 --> 00:05:07,320 editors in chief of Computers & Fluids. 93 00:05:08,000 --> 00:05:11,280 And he gave us a book by Shapiro. 94 00:05:12,000 --> 00:05:14,720 It's fast profiles. I think it's a book. 95 00:05:14,720 --> 00:05:16,680 You know, Shapiro is a, is a professor. 96 00:05:16,680 --> 00:05:18,696 He was professor at MIT I think. 97 00:05:18,704 --> 00:05:22,980 And he said he he gave us the book the first day of the 98 00:05:22,988 --> 00:05:27,186 classroom and he said if you don't like this book, you can 99 00:05:27,194 --> 00:05:30,396 change, go to another take another course to just abandoned 100 00:05:30,396 --> 00:05:34,181 the fluid mechanics option. And I read it and I loved it so. 101 00:05:34,181 --> 00:05:35,840 And that's how I meant for the 102 00:05:35,840 --> 00:05:40,400 mechanics. So that's interesting. 103 00:05:40,400 --> 00:05:45,313 And then did you always want to go and do a PhD or were you 104 00:05:45,313 --> 00:05:48,240 sort of debating of going into industry or something? 105 00:05:48,560 --> 00:05:51,240 Yes, because you know my family, everybody. 106 00:05:51,680 --> 00:05:53,760 So my family is a family of professors. 107 00:05:53,760 --> 00:05:57,480 Well, they are professors more in the secondary school, but 108 00:05:57,480 --> 00:06:02,120 they all did high studies, let's say university and so on. 109 00:06:02,680 --> 00:06:05,720 So I wanted to be a professor and I wanted to be a researcher 110 00:06:05,720 --> 00:06:08,440 most first of all, in the 1st place, I wanted to be a 111 00:06:08,440 --> 00:06:14,200 researcher And, and So what I knew that was necessary to have 112 00:06:14,200 --> 00:06:19,320 a PhD for being a researcher. And so well, since the beginning 113 00:06:19,320 --> 00:06:23,360 I was looking for a PhD. And at that time it was very 114 00:06:23,360 --> 00:06:27,280 hard to have a PhD in Italy because there was no traditional 115 00:06:27,280 --> 00:06:33,960 PhD, I would say 30 years ago. And the people went abroad for 116 00:06:33,960 --> 00:06:37,840 the PhD actually, because in Italy it was not very valued. 117 00:06:38,320 --> 00:06:42,200 It was at the very beginning. I was maybe the 11th cycle of 118 00:06:42,200 --> 00:06:46,680 PhD, which means that the, the PhD degree in Italy was created 119 00:06:46,680 --> 00:06:50,120 11 years before I start, you know, so it was relatively 120 00:06:50,320 --> 00:06:56,930 recent and, and there were very few fellowships for a PhD, so 121 00:06:56,930 --> 00:06:59,622 it was very difficult to have them. 122 00:06:59,630 --> 00:07:04,420 And that's how at some point I, I decided to move to France. 123 00:07:04,428 --> 00:07:07,208 So, yeah. And and then, well, I loved 124 00:07:07,216 --> 00:07:10,950 France and remained here. Yeah, OK. 125 00:07:10,958 --> 00:07:16,000 And your topic was on more numerical schemes, more like 126 00:07:16,008 --> 00:07:21,116 compressible to how did you get to that, you know, decision I I 127 00:07:21,116 --> 00:07:24,072 guess on the. Topic Actually the very 128 00:07:24,080 --> 00:07:26,728 beginning. At the very beginning I was 129 00:07:26,736 --> 00:07:28,500 doing schemes for incompressible flow. 130 00:07:28,508 --> 00:07:33,090 It was I started with the lid driven cavity and the velocity 131 00:07:33,098 --> 00:07:35,916 vorticity formulation of the Navier–Stokes equation. 132 00:07:35,924 --> 00:07:40,968 So that was my master's thesis, but then when I was looking for 133 00:07:40,968 --> 00:07:44,904 a PhD for a PhD in France, actually the beginning should 134 00:07:44,904 --> 00:07:47,480 have been just a second master degree. 135 00:07:47,560 --> 00:07:50,440 And after I decided to remain for the PhD in France. 136 00:07:50,920 --> 00:07:55,760 But anyway, so my, my supervisor in Italy knew Professor Alain 137 00:07:55,760 --> 00:07:59,634 Lerat, who was one of the founders of the ICCFD 138 00:07:59,634 --> 00:08:02,483 conference too. And he said, well, I know this 139 00:08:02,491 --> 00:08:05,240 guy in France is doing good job with, but it's compressible. 140 00:08:05,248 --> 00:08:08,448 So it's another thing. And then if you don't, you're 141 00:08:08,456 --> 00:08:11,388 not scared about compressible flows, you can go to him and 142 00:08:11,396 --> 00:08:13,544 he's very good. She's doing, she's doing 143 00:08:13,552 --> 00:08:17,696 numerical schemes and so on. And so that's how I decided, oh, 144 00:08:17,696 --> 00:08:20,395 Paris is not bad. Actually it was because of Paris 145 00:08:20,395 --> 00:08:22,992 more than because of the compressible schemes, because my 146 00:08:22,992 --> 00:08:27,382 options were to go to the von Kármán Institute, to go to the 147 00:08:27,382 --> 00:08:31,709 US or to go to Paris. And what I said about Paris, I 148 00:08:31,717 --> 00:08:36,438 have to go to Paris. And so I arrived here and 149 00:08:36,446 --> 00:08:40,850 Alain Lerat proposed to me to work on high order schemes which 150 00:08:40,850 --> 00:08:43,840 were a quite recent topic at that time. 151 00:08:43,847 --> 00:08:48,108 You know, at that time high order was second order schemes 152 00:08:48,116 --> 00:08:50,534 actually because everybody was doing 1st order. 153 00:08:50,542 --> 00:08:52,872 So 2nd order scheme were already high order. 154 00:08:52,880 --> 00:08:56,650 And he said to me, but we are going to move to 3rd order. 155 00:08:56,658 --> 00:08:59,030 OK, yeah. And that's, you do unsteady 156 00:08:59,038 --> 00:09:02,760 flows because at the, you know, at that time people were mostly 157 00:09:02,760 --> 00:09:05,005 doing steady Euler or steady Navier–Stokes. 158 00:09:05,013 --> 00:09:10,100 And so, well, I started a PhD for high order scheme to capture 159 00:09:10,100 --> 00:09:13,880 steady phenomena, even if it was only RANS at the time. 160 00:09:13,888 --> 00:09:17,558 And, and, and yes, it was a high order finite-volume schemes. 161 00:09:17,566 --> 00:09:21,744 And then I spent quite a lot of my career on on high order 162 00:09:21,752 --> 00:09:26,688 scheme because of that. And, and was it therefore a 163 00:09:26,696 --> 00:09:30,361 natural progression to continue going down the route to sort of 164 00:09:30,361 --> 00:09:33,840 post doc and, and, and you know, faculty position with was 165 00:09:33,840 --> 00:09:40,888 it just an evolution or, or was it a, a challenge to, to, to 166 00:09:40,888 --> 00:09:45,942 progress down that route? It was a challenge because as 167 00:09:45,950 --> 00:09:49,721 you know, there are not many positions in academia. 168 00:09:49,729 --> 00:09:56,554 They're very challenging. Also, when I, when I was in 169 00:09:56,562 --> 00:10:01,275 France, I spent quite one year, the second master degree and 170 00:10:01,283 --> 00:10:03,374 then three years of PhD in France. 171 00:10:03,382 --> 00:10:07,162 And at some point I wanted to go back to Italy. 172 00:10:07,170 --> 00:10:11,220 And in Italy there were zero positions basically. 173 00:10:11,228 --> 00:10:18,100 So I passed the competitions in France because in France, you 174 00:10:18,108 --> 00:10:21,512 know, the positions are open, you have to to get a 175 00:10:21,520 --> 00:10:23,670 qualification, which is a national qualification. 176 00:10:23,678 --> 00:10:28,000 So this depends on how many papers if you have taught 177 00:10:28,008 --> 00:10:33,295 courses not and so on. And you it's basically sort of 178 00:10:33,303 --> 00:10:39,000 minimum certification, which says, OK, you are fit to be a 179 00:10:39,008 --> 00:10:41,970 professor. OK, not, not a professor, but an 180 00:10:41,970 --> 00:10:45,681 assistant professor. And then we once you have got 181 00:10:45,689 --> 00:10:50,080 got, once you got the, the, the qualification, you have to apply 182 00:10:50,080 --> 00:10:53,728 at the universities which have positions open and you have to 183 00:10:53,736 --> 00:10:57,554 compete with other guys. So I eventually got a position 184 00:10:57,562 --> 00:11:02,504 here in Paris, but I wanted to go back to Italy. 185 00:11:02,512 --> 00:11:06,676 So I renounced the position, even it was rank at first. 186 00:11:06,684 --> 00:11:09,345 And I decided to go back to Italy. 187 00:11:09,353 --> 00:11:13,810 And there I started with a post doc because there were no 188 00:11:13,818 --> 00:11:18,298 permanent positions open. And after some time after my 189 00:11:18,306 --> 00:11:23,420 post doc, I moved to a close by university because, well, my 190 00:11:23,428 --> 00:11:27,495 home university is Bari. It's in the South of Italy, it's 191 00:11:27,495 --> 00:11:30,223 Apulia. And I went to Lecce, which is 192 00:11:30,231 --> 00:11:35,880 even even more south you know, it's the very tip of the heel. 193 00:11:35,888 --> 00:11:40,728 And, and there I got the eventually a faculty position 194 00:11:40,736 --> 00:11:45,540 and I worked there for eight years as an assistant professor. 195 00:11:45,540 --> 00:11:48,200 And I was the only assistant 196 00:11:48,200 --> 00:11:51,120 professor in fluid dynamics of the whole university. 197 00:11:51,240 --> 00:11:58,440 I was the only one. And, and what was what was your 198 00:11:58,440 --> 00:12:02,640 main focus back was back then, you know, when did this move to 199 00:12:02,640 --> 00:12:06,600 the the RANS modelling and certainty quantification and 200 00:12:06,600 --> 00:12:08,840 things like that? Was that during that period or 201 00:12:08,840 --> 00:12:11,000 was that? Actually when I arrived in 202 00:12:11,000 --> 00:12:16,480 Lecce, I was, as I said, I was the only fluid mechanician and I 203 00:12:16,480 --> 00:12:21,400 wasn't part of a group, a larger group in energetics. 204 00:12:22,280 --> 00:12:26,560 And I started to think about what could I do which is related 205 00:12:26,560 --> 00:12:31,240 to energetics. And I discovered almost by by by 206 00:12:31,240 --> 00:12:35,800 chance dense gases. And so dense gases are 207 00:12:35,800 --> 00:12:39,760 compressible. So it's it's gases. 208 00:12:39,760 --> 00:12:43,560 So it's compressible. Compressible flows of organic 209 00:12:43,560 --> 00:12:47,400 fluids, which are governed by complex equations of state could 210 00:12:47,400 --> 00:12:50,640 be super critical CO2, but it's it's light gas. 211 00:12:51,040 --> 00:12:55,000 But you can have denser gases or more heavier gases like 212 00:12:55,400 --> 00:13:01,080 refrigerants or hydrocarbons. And these gases are used in 213 00:13:01,080 --> 00:13:06,160 processes, industrial processes, or also in some thermodynamic 214 00:13:06,160 --> 00:13:09,920 cycles like, well, of course the refrigeration cycles, but also 215 00:13:11,040 --> 00:13:14,880 direct cycles, which means, for instance, organic Rankine 216 00:13:14,880 --> 00:13:17,920 cycles, which are like the Rankine cycle, which works with 217 00:13:17,920 --> 00:13:20,400 these strange gases instead of water. 218 00:13:21,360 --> 00:13:25,687 And the problem with these gases is that since it is 219 00:13:25,687 --> 00:13:29,486 industrial, industrial fluids they are very ill 220 00:13:29,486 --> 00:13:32,264 characterized. Actually you don't know exactly 221 00:13:32,264 --> 00:13:35,455 about the properties about equational state you should use 222 00:13:35,455 --> 00:13:40,196 there are few data. The material properties are 223 00:13:40,204 --> 00:13:45,045 given in only technical sheets with a lot of uncertainties. 224 00:13:45,053 --> 00:13:47,446 And that's how I came to uncertainty quantification, 225 00:13:47,454 --> 00:13:50,796 because at some point I discovered polynomial chaos 226 00:13:50,796 --> 00:13:54,772 stuff and so on. And I said, oh, this is perfect 227 00:13:54,772 --> 00:13:58,286 because I have a fluid where actually don't know exactly how 228 00:13:58,286 --> 00:14:01,440 it behaves. So I would like to characterize 229 00:14:01,440 --> 00:14:06,304 the impact on the CFD solution of the, the, the, the bad 230 00:14:06,312 --> 00:14:09,220 knowledge of the fluid properties, not only the 231 00:14:09,228 --> 00:14:12,064 thermodynamics, but also the transport properties are not 232 00:14:12,072 --> 00:14:16,148 very well known. And so, yeah, that's how I moved 233 00:14:16,148 --> 00:14:19,142 to uncertainty quantification. But at the beginning was not 234 00:14:19,150 --> 00:14:24,296 RANS, it was even Euler, you know, but you don't know exactly 235 00:14:24,296 --> 00:14:27,624 the thermodynamic, you cannot characterize completely the 236 00:14:27,632 --> 00:14:31,402 thermodynamic behavior of your gas because you, you, you don't 237 00:14:31,402 --> 00:14:34,446 know which equation of state you should use. 238 00:14:34,454 --> 00:14:38,720 And the parameters of the equation of state are, you know, 239 00:14:38,720 --> 00:14:43,130 not very accurate. That's interesting. 240 00:14:43,138 --> 00:14:48,452 So it's more driven from that. Yeah, I guess did that help you 241 00:14:48,452 --> 00:14:51,890 working in a more complex area in a way? 242 00:14:51,898 --> 00:14:53,909 Did it give you more understanding of the 243 00:14:53,917 --> 00:14:57,601 fundamentals of fluid mechanics? I guess when people just study 244 00:14:57,609 --> 00:15:00,280 single phase simple incompressible, maybe they don't 245 00:15:00,280 --> 00:15:02,290 appreciate the. Complexity, yeah. 246 00:15:02,298 --> 00:15:07,720 Actually I I adore those gases because, well at least 247 00:15:07,728 --> 00:15:11,200 theoretically they could exhibit very exotic behaviors. 248 00:15:11,208 --> 00:15:15,294 In particular they are expected to exhibit under some 249 00:15:15,302 --> 00:15:18,697 thermodynamic conditions. If if you take a heavy enough 250 00:15:18,705 --> 00:15:21,880 gas, it is expected to exhibit expansion shock waves. 251 00:15:21,888 --> 00:15:26,236 And actually there is a community around that which is 252 00:15:26,244 --> 00:15:30,018 called the non ideal compressible fluid dynamics 253 00:15:30,026 --> 00:15:34,075 community, which has spent several years trying to expect 254 00:15:34,083 --> 00:15:38,720 to to have an experimental proof of the existence of these 255 00:15:38,728 --> 00:15:42,002 expansion shock waves in single phase compressible flows. 256 00:15:42,010 --> 00:15:46,064 So the first one to postulate that is Hans Bethe. 257 00:15:46,072 --> 00:15:49,804 So the the physician, the physicist Hans Bethe, Nobel 258 00:15:49,812 --> 00:15:56,946 Prize who showed that for van der Waals gases with some values 259 00:15:56,946 --> 00:16:03,480 of the of the polytropic exponent, you can get a region 260 00:16:03,480 --> 00:16:08,469 where the second principle of thermodynamics forbids the the 261 00:16:08,477 --> 00:16:12,292 classical shock waves. So the the the compression shock 262 00:16:12,292 --> 00:16:15,440 waves. So instead you have a 263 00:16:15,448 --> 00:16:19,480 compression fan and conversely, you can have expansion shock 264 00:16:19,488 --> 00:16:21,832 waves and instead of expansion fan. 265 00:16:21,840 --> 00:16:25,800 So everything is reversed. And then several researchers 266 00:16:25,808 --> 00:16:30,020 work on that. Thompson, Michael Cramer at at 267 00:16:30,028 --> 00:16:35,268 Virginia Tech and so on. And people were fascinated by 268 00:16:35,276 --> 00:16:38,740 these gases because everybody wanted to prove experimentally 269 00:16:38,748 --> 00:16:42,328 that that was possible. And eventually there was a group 270 00:16:42,328 --> 00:16:46,740 in Russia at some point we did an experiment with FC-70, which 271 00:16:46,740 --> 00:16:50,864 is a, you know, fluorocarbon, those which are very, very bad 272 00:16:50,872 --> 00:16:53,962 for the ozone layer. So they did an experiment with 273 00:16:53,970 --> 00:16:57,140 this and they said, oh, here, here you have the expansion 274 00:16:57,148 --> 00:17:00,128 shockwaves. But then the people started to 275 00:17:00,136 --> 00:17:03,300 challenge the experiment and said no, this is impossible, 276 00:17:03,308 --> 00:17:05,545 Probably it was two-phase and so on. 277 00:17:05,553 --> 00:17:08,006 And then in depth there was a group. 278 00:17:08,013 --> 00:17:13,105 So the group of Piero Colonna, there is a big group in TU Delft 279 00:17:13,105 --> 00:17:16,592 who tried to reproduce an experiment with different gases. 280 00:17:16,592 --> 00:17:18,680 Also you have Alberto Guardone, 281 00:17:18,680 --> 00:17:22,000 in Milan. He, he got an ERC actually on 282 00:17:22,000 --> 00:17:24,560 that. And what it's very, very 283 00:17:24,560 --> 00:17:28,240 difficult because this shock expansion shock waves exist in a 284 00:17:28,240 --> 00:17:31,640 very tiny thermodynamic region and all the uncertainties 285 00:17:31,640 --> 00:17:35,440 associated with the shock tube with the membrane breaking and 286 00:17:35,440 --> 00:17:39,320 so on can perturb the actual development of the shockwave. 287 00:17:39,320 --> 00:17:43,440 And so we are never sure it's really a pure expansion 288 00:17:43,440 --> 00:17:47,960 shockwave. So, well, and the interest of 289 00:17:47,960 --> 00:17:50,800 that thing besides the, let's say the, the, the, the 290 00:17:50,800 --> 00:17:55,800 scientific curiosity was that some people were expecting that 291 00:17:55,920 --> 00:18:00,852 you could exploit this behaviour to get better energy 292 00:18:00,852 --> 00:18:04,280 conversion cycle with the reduced losses and so on. 293 00:18:05,600 --> 00:18:10,640 It has been abandoned right now a little bit because actually 294 00:18:11,600 --> 00:18:14,440 while there is a very small industry supporting that because 295 00:18:14,560 --> 00:18:19,000 it's very specific machines and it's very small companies 296 00:18:19,000 --> 00:18:22,880 working on that with few funds for to invest in research. 297 00:18:23,560 --> 00:18:28,080 But it was a great period. So it was actually, and also it 298 00:18:28,080 --> 00:18:30,800 has the impact, even if you don't find the expansion shock 299 00:18:30,800 --> 00:18:33,110 waves, at some point, you don't care because there are plenty 300 00:18:33,110 --> 00:18:36,276 of systems with real gas effects, not expansion 301 00:18:36,276 --> 00:18:38,420 shockwave, but still real gas effects. 302 00:18:38,428 --> 00:18:42,120 And that you need to characterize to have better 303 00:18:42,128 --> 00:18:46,778 conversion cycles or better heat pumps or, you know, all these 304 00:18:46,786 --> 00:18:48,719 stuff. And so we're still working. 305 00:18:48,727 --> 00:18:53,180 So we are right now a project ongoing with the, with, with the 306 00:18:53,180 --> 00:18:56,116 German team. We are doing the, the 307 00:18:56,124 --> 00:18:59,320 simulations and the machine learning and they are doing the 308 00:18:59,320 --> 00:19:01,528 experiments and it's a very nice team. 309 00:19:01,536 --> 00:19:05,280 And yeah, we, yeah. Very cool. 310 00:19:05,288 --> 00:19:10,344 And you, you mentioned at the beginning that, you know, you 311 00:19:10,352 --> 00:19:14,664 were in Italy, you went to Paris, you went back to Italy. 312 00:19:14,672 --> 00:19:18,948 What then was the what? What led to you, you know, going 313 00:19:18,948 --> 00:19:23,859 back to Paris again? Yeah, well, again, I mean, it's 314 00:19:23,859 --> 00:19:29,856 not easy to stay in academia in Italy because the, you know, the 315 00:19:29,856 --> 00:19:34,885 funding system, well, the academic system in Italy is 316 00:19:34,893 --> 00:19:37,195 chronically underfunded, but severely underfunded. 317 00:19:37,203 --> 00:19:39,320 That's why you find Italians everywhere. 318 00:19:39,328 --> 00:19:43,035 Actually, I've noticed maybe that there are lots of Italians 319 00:19:43,035 --> 00:19:47,215 everywhere. So, so the, the academic 320 00:19:47,223 --> 00:19:50,290 position are very few. I had one actually. 321 00:19:50,298 --> 00:19:54,314 But then you are alone. It's very difficult to get 322 00:19:54,322 --> 00:19:57,860 funding. The possibility of having a 323 00:19:57,868 --> 00:20:02,423 career are very, So you, you have to be very patient, you 324 00:20:02,431 --> 00:20:06,784 know. And also, my husband wasn't 325 00:20:06,792 --> 00:20:12,325 Italian and he never could find a satisfactory job in Italy. 326 00:20:12,333 --> 00:20:16,346 So at some point he said, well, let's go back to Paris. 327 00:20:16,354 --> 00:20:20,247 And there I applied because I wanted to move to a 328 00:20:20,255 --> 00:20:22,229 professorship. I was assistant professor, I 329 00:20:22,237 --> 00:20:26,229 wanted to move a professor and I got a position in Paris. 330 00:20:26,237 --> 00:20:29,911 And that's how we we moved back to Paris because he also had a 331 00:20:29,919 --> 00:20:34,594 job here in Paris. Ah, OK, I understand. 332 00:20:34,602 --> 00:20:40,716 So you well, as you say, Paris is not such a bad place to to to 333 00:20:40,716 --> 00:20:43,248 be. Easy for two person It's easier 334 00:20:43,248 --> 00:20:47,872 to find both a job in Paris than find both a job in the in 335 00:20:47,880 --> 00:20:52,857 southern Italy. Yeah, but that does seem to be, 336 00:20:52,857 --> 00:20:58,570 I guess, an overriding thing of academia that it's almost this 337 00:20:58,578 --> 00:21:02,218 necessity to move, right. It's it's kind of a weird 338 00:21:02,226 --> 00:21:04,730 profession. Even in the US or the UK, very 339 00:21:04,738 --> 00:21:08,185 few people will have their entire career at one place, 340 00:21:08,193 --> 00:21:12,594 particularly earlier on. They will have to have you found 341 00:21:12,594 --> 00:21:16,450 that it's almost, it's a fairly unfair thing in a way 342 00:21:16,450 --> 00:21:20,067 that you have to move your family and do things just to 343 00:21:20,067 --> 00:21:22,520 progress. Yeah, yeah, that that that's 344 00:21:22,520 --> 00:21:26,760 true. And that's because, well, for 345 00:21:26,760 --> 00:21:29,600 instance, in France, there is even a rule in some, in some 346 00:21:29,600 --> 00:21:32,120 disciplines, like in mathematics, that if you have 347 00:21:32,120 --> 00:21:36,520 been an assistant professor in a department, you cannot be a 348 00:21:37,280 --> 00:21:39,000 professor in the same department. 349 00:21:39,000 --> 00:21:43,040 You have to move another one. And these small cities where you 350 00:21:43,040 --> 00:21:46,920 have just one university, it's basically this means that you 351 00:21:46,920 --> 00:21:49,600 have to move. So this pushes a lot of people 352 00:21:49,600 --> 00:21:53,377 to abandon the idea of moving to a professorship and they 353 00:21:53,377 --> 00:21:56,240 remain assistant professors for forever. 354 00:21:56,600 --> 00:21:58,600 Also because the French system allows it. 355 00:21:58,600 --> 00:22:00,880 It's not like the tenure track in the US. 356 00:22:00,880 --> 00:22:04,160 If you don't, if you are not promoted to associate, then you 357 00:22:04,160 --> 00:22:08,280 have to leave academia. You can remain as associate 358 00:22:08,280 --> 00:22:10,880 professor with a permanent position forever. 359 00:22:11,240 --> 00:22:16,280 So if you can't move with your family, basically you cannot 360 00:22:17,280 --> 00:22:22,120 progress in your career. And and so it's a choice. 361 00:22:23,720 --> 00:22:27,080 Some people are happy with the, with an assistant professorship 362 00:22:27,080 --> 00:22:29,400 for their life, they do teaching and so on. 363 00:22:30,240 --> 00:22:33,974 But if you, if you want to move, so if you want to progress 364 00:22:33,974 --> 00:22:37,080 with the career, yes. The, the, I don't know, the 365 00:22:37,080 --> 00:22:41,400 academic, the academic world is built like that. 366 00:22:41,400 --> 00:22:45,680 Why don't know exactly it's also part of science to move. 367 00:22:45,680 --> 00:22:49,240 We should look in the past at the the former scientists where 368 00:22:49,240 --> 00:22:53,280 they were moving quite a lot. You know, even in the 369 00:22:53,280 --> 00:22:56,720 Renaissance, you know, you, you, you had this, I don't know 370 00:22:56,720 --> 00:23:00,680 Galileo or or Leonardo da Vinci, they were moving around. 371 00:23:01,320 --> 00:23:06,560 And that's because you need to to spread knowledge to find 372 00:23:06,560 --> 00:23:09,440 better environment. So I think it's nice the 373 00:23:09,440 --> 00:23:14,640 difficulties when you also have a family life and you have to, 374 00:23:14,880 --> 00:23:17,040 so you have two person to, to, to move. 375 00:23:17,920 --> 00:23:22,400 In some cases it's easy because 1 is a flexible job or, or 376 00:23:22,600 --> 00:23:25,960 things at home. But otherwise it can be, yeah, 377 00:23:26,720 --> 00:23:29,280 quite challenging. And I think it's something that 378 00:23:29,280 --> 00:23:33,360 the new generation don't appreciate that much because, 379 00:23:33,400 --> 00:23:38,400 you know, I, I think I am what's, what's my generation, I 380 00:23:38,400 --> 00:23:43,240 think generation X and we are, you know, I, I know that my 381 00:23:43,240 --> 00:23:46,520 children say your generation X is the generation of suffering. 382 00:23:46,600 --> 00:23:51,120 You know you were you were raised to suffer, but it's no 383 00:23:51,120 --> 00:23:55,880 longer the case. But it, but it is a serious 384 00:23:55,880 --> 00:23:59,440 point though, that I've, I almost feel like sometimes 385 00:23:59,440 --> 00:24:06,800 academia has maybe not to be careful, but I always, I'm just 386 00:24:06,800 --> 00:24:11,840 amazed that the dedication it requires to get through it and 387 00:24:11,840 --> 00:24:14,520 to to become Someone Like You, you know, a top professor. 388 00:24:14,520 --> 00:24:21,267 Like it feels, it's, I feel like probably many people aren't 389 00:24:21,267 --> 00:24:26,920 able to get to that .1 Obviously they're not intellectual 390 00:24:26,920 --> 00:24:28,614 enough. They're not, you know, capable 391 00:24:28,622 --> 00:24:31,656 enough. But but also you have to have a 392 00:24:31,656 --> 00:24:34,515 quiet determination. Yeah, to, to get there, which is 393 00:24:34,515 --> 00:24:37,520 probably partly a good thing because it, you know, it's, it's 394 00:24:37,520 --> 00:24:40,464 like a, you know, survival of the fittest, I guess. 395 00:24:40,472 --> 00:24:44,450 But as you've said to some people where they just really 396 00:24:44,458 --> 00:24:48,012 would love to stay where they're living because they have friends 397 00:24:48,012 --> 00:24:51,532 and family and they can't, it's, it's a shame. 398 00:24:51,540 --> 00:24:55,480 Well, what I used to say to my PhD students is that the 399 00:24:55,488 --> 00:25:01,204 academic career is kind of similar to a career in music or 400 00:25:01,204 --> 00:25:04,898 in theatre or in high level level sport. 401 00:25:04,906 --> 00:25:09,419 OK, if you are a musician, if you're a pianist or if you are 402 00:25:09,427 --> 00:25:13,480 an opera singer, you have to move around and it's a very 403 00:25:13,488 --> 00:25:16,478 competitive field. And you do that not because you 404 00:25:16,478 --> 00:25:19,960 want no quiet life, a quiet life, but because you are 405 00:25:19,968 --> 00:25:24,792 passioned and you want to be on the stage and see, you know, 406 00:25:24,800 --> 00:25:26,880 people around you clapping hands. 407 00:25:26,888 --> 00:25:32,262 And, you know, it's a sort of, you know, I think that well, 408 00:25:32,270 --> 00:25:36,716 there is a sort of narcissistic side, maybe like people 409 00:25:36,724 --> 00:25:41,384 remaining in academia, but also, you know, it's it's an 410 00:25:41,392 --> 00:25:45,764 intellectual work. And I guess that what if you, 411 00:25:45,772 --> 00:25:49,960 you should you, you do that because because of passion, 412 00:25:49,968 --> 00:25:54,384 first of all, it's not, it's not an ordinary job actually. 413 00:25:54,392 --> 00:26:00,456 And you want to push the, the, the, the, the, the, the, the, 414 00:26:00,464 --> 00:26:04,961 the frontiers of knowledge. You want your, your work to have 415 00:26:04,961 --> 00:26:09,355 an impact and you know, you want to do something to move 416 00:26:09,355 --> 00:26:12,600 things forward. And, and that's a, that's a 417 00:26:12,608 --> 00:26:16,036 passion. So it's maybe you don't, you 418 00:26:16,044 --> 00:26:21,015 don't make it, but you still try because you know, it's, it's 419 00:26:21,023 --> 00:26:24,122 like a football player. Not not all football players are 420 00:26:24,122 --> 00:26:27,868 Lionel Messi or, I don't know, David Beckham. 421 00:26:27,876 --> 00:26:33,142 That's a very good. I've not heard it described that 422 00:26:33,142 --> 00:26:35,600 way about a musician or sports players. 423 00:26:35,608 --> 00:26:38,528 It's kind of true that they also move around a lot. 424 00:26:38,536 --> 00:26:43,230 And I, I get that's the, that's the, the old thing though of 425 00:26:43,238 --> 00:26:46,162 universities, isn't it, that you almost have to wait for the 426 00:26:46,170 --> 00:26:49,412 person above you to, you know, retire or die to get their 427 00:26:49,420 --> 00:26:51,515 position. So people end up also moving 428 00:26:51,523 --> 00:26:55,360 around because there's sort of not enough where I guess in a 429 00:26:55,368 --> 00:26:58,610 big corporate world, people are more able to, maybe that's 430 00:26:58,618 --> 00:27:02,250 changing, But you know, in a, in like an Airbus or something, 431 00:27:02,258 --> 00:27:07,050 I'm, I'm sure there's jobs for life almost just in one company 432 00:27:07,050 --> 00:27:10,576 where academia. So I, I, I say that because I 433 00:27:10,584 --> 00:27:14,200 think it, people should appreciate just how hard it is 434 00:27:14,208 --> 00:27:18,070 to get to being a professor, that it, it's not just an 435 00:27:18,078 --> 00:27:20,725 intellectual thing. It's like a determination and 436 00:27:20,733 --> 00:27:25,440 passion thing, as you say, to, to, to do it. 437 00:27:25,448 --> 00:27:30,028 And so all, you know, talking about, you know, wanting to be 438 00:27:30,036 --> 00:27:33,742 at the forefront. I mean, you, you were quite 439 00:27:33,750 --> 00:27:36,789 early on looking at learning and data-driven turbulence models 440 00:27:36,789 --> 00:27:40,360 and and machine learning. You know, when did you start to 441 00:27:40,360 --> 00:27:43,551 get into that? When did you get a sense that 442 00:27:43,559 --> 00:27:47,272 these methods were an alternative or an enhancement on 443 00:27:47,272 --> 00:27:50,687 the more traditional numerical methods sort of line of 444 00:27:50,695 --> 00:27:53,224 research? Well, actually I. 445 00:27:53,232 --> 00:27:57,175 Stepped into Bayesian methods when I was doing uncertainty 446 00:27:57,183 --> 00:28:02,120 quantification for dense gases. And at some point I was 447 00:28:02,128 --> 00:28:06,004 discussing with some colleagues of mine from from the University 448 00:28:06,004 --> 00:28:10,520 of Trieste, and they had a sort of startup, I don't know if 449 00:28:10,520 --> 00:28:12,560 you know this startup called ESTECO. 450 00:28:12,640 --> 00:28:17,480 They produce software called modeFRONTIER. 451 00:28:18,080 --> 00:28:20,440 Oh yes. Yes, yes, of course, yeah. 452 00:28:21,400 --> 00:28:25,160 And so, yeah, they were actually promoting this software in 453 00:28:25,160 --> 00:28:31,000 universities and, and they sold also academic licenses and so 454 00:28:31,000 --> 00:28:33,760 on. And so I started to discuss with 455 00:28:33,760 --> 00:28:36,960 them about my problems with uncertainty quantification in 456 00:28:36,960 --> 00:28:40,480 dense gases. And they said, OK, but you could 457 00:28:40,480 --> 00:28:42,840 try to solve an inverse problem and so on. 458 00:28:42,840 --> 00:28:47,440 And, and, and so that I, so I started to look into the 459 00:28:47,440 --> 00:28:52,320 literature and I started to be interested into these Bayesian 460 00:28:52,320 --> 00:28:58,840 methods and, and I said, OK, this could work for dense gases, 461 00:28:58,840 --> 00:29:01,680 but we, well, the thermodynamics is not the only source of 462 00:29:01,680 --> 00:29:04,360 uncertainty because you also have the uncertainties 463 00:29:04,360 --> 00:29:05,960 associated with the turbulence models. 464 00:29:05,960 --> 00:29:07,880 And these are very old problems. 465 00:29:07,888 --> 00:29:11,262 So there are so many turbulence models, you don't know which one 466 00:29:11,262 --> 00:29:14,085 you have to choose, you don't know which parameters you should 467 00:29:14,085 --> 00:29:18,360 put into them and so on. So let's try to quantify this 468 00:29:18,368 --> 00:29:22,680 statistically instead of just using expert knowledge, which is 469 00:29:22,680 --> 00:29:27,064 what any anybody does actually, because every, every 470 00:29:27,064 --> 00:29:33,982 company actually has sort of, you know, established know-how 471 00:29:33,982 --> 00:29:38,524 saying OK, for this problem, you should use the k-ω SST for 472 00:29:38,524 --> 00:29:41,670 this problem, you should use the Spalart–Allmaras for this 473 00:29:41,670 --> 00:29:42,960 problem, you should use that one. 474 00:29:43,160 --> 00:29:46,440 And sometimes you also know that they retune the parameters. 475 00:29:46,800 --> 00:29:51,320 For instance, I know that people in internal combustion engine 476 00:29:51,320 --> 00:29:57,227 used to recalibrate k-ε for for having better results for, 477 00:29:57,227 --> 00:30:00,320 for, for, for, Yeah, for internal combustion engines. 478 00:30:00,760 --> 00:30:04,240 And but all this was was tuned by hands basically. 479 00:30:04,640 --> 00:30:08,840 And I said, OK, if there are, you know, mathematical, clean 480 00:30:08,840 --> 00:30:12,240 mathematical techniques to do that, let's try to use them. 481 00:30:12,240 --> 00:30:16,040 And so since I was already doing direct uncertainty 482 00:30:16,040 --> 00:30:20,400 quantification, I started to, you know, try to do the the 483 00:30:20,400 --> 00:30:22,440 inverse uncertainty quantification. 484 00:30:23,320 --> 00:30:26,680 And at some point, well, it was a bit hard for me because you 485 00:30:26,680 --> 00:30:30,171 know, in mechanical engineering we are not very well trained 486 00:30:30,171 --> 00:30:33,148 in, in probability and statistics. 487 00:30:33,156 --> 00:30:37,125 So eventually. I, I studied that myself because 488 00:30:37,125 --> 00:30:41,652 it was interesting to that actually I also thought a little 489 00:30:41,652 --> 00:30:45,767 bit, but it was not enough to, you know, understand all the 490 00:30:45,775 --> 00:30:49,448 details. And so at some point I was 491 00:30:49,456 --> 00:30:53,500 invited by statistics department in Chile and they invited me for 492 00:30:53,500 --> 00:30:57,400 two months and I was there with statisticians and 493 00:30:57,400 --> 00:30:59,680 mathematicians of probabilities. 494 00:30:59,680 --> 00:31:05,000 So and during the first month it was impossible to understand 495 00:31:05,000 --> 00:31:08,920 each other because, you know, I was calling things with names 496 00:31:08,920 --> 00:31:11,240 that they interpreted in another way and so on. 497 00:31:12,240 --> 00:31:15,480 And, but at some point we started to understand each other 498 00:31:15,480 --> 00:31:19,480 and that was great because some papers which were, you know, 499 00:31:19,480 --> 00:31:26,600 just hieroglyphs to me, hieroglyphs started to be clear. 500 00:31:26,720 --> 00:31:31,340 And, and that's where, you know, all the, the Bayesian stuff 501 00:31:31,340 --> 00:31:35,640 was, was set and, and, and I could start to recalibrate the 502 00:31:35,640 --> 00:31:37,320 turbulence models and and so on. 503 00:31:37,328 --> 00:31:41,368 So that. So yeah, the fact of being in an 504 00:31:41,368 --> 00:31:44,264 interdisciplinary environment and talk with mathematicians was 505 00:31:44,264 --> 00:31:48,340 extremely useful to me, not only with engineers, even if I 506 00:31:48,340 --> 00:31:52,085 love engineers, but you know, sometimes like, no, that. 507 00:31:52,093 --> 00:31:54,406 That's, that's a very good point. 508 00:31:54,414 --> 00:31:59,380 And I guess even to today, that's a challenge on the 509 00:31:59,388 --> 00:32:03,140 machine learning side, isn't it, that maybe it's starting to 510 00:32:03,148 --> 00:32:07,708 change in a course today. But I guess most people who are 511 00:32:07,708 --> 00:32:10,996 doing the research, therefore who studied, you know, 5 to 10 512 00:32:11,004 --> 00:32:15,130 years ago or more didn't have any of that background in 513 00:32:15,138 --> 00:32:18,660 computer science or in statistics or, or sort of 514 00:32:18,668 --> 00:32:21,760 methods that maybe people who did more maths, stronger maths 515 00:32:21,768 --> 00:32:26,400 may, may, may do it, but engineering courses probably 516 00:32:26,408 --> 00:32:29,136 wouldn't. I mean, how have you seen that 517 00:32:29,144 --> 00:32:30,434 affect the machine learning side? 518 00:32:30,442 --> 00:32:32,588 I know you've been quite passionate about connecting. 519 00:32:32,596 --> 00:32:38,385 I know you you kicked off, for example, with the Extrality and 520 00:32:38,385 --> 00:32:42,034 AirfRANS data sets. And, you know, did you try and 521 00:32:42,042 --> 00:32:45,220 take some of that inspiration of your time getting familiar with 522 00:32:45,220 --> 00:32:48,894 statistics and try and do the same on the machine learning 523 00:32:48,902 --> 00:32:50,641 side? Yeah, yeah, definitely. 524 00:32:50,649 --> 00:32:55,050 Actually, I had the chance when, when I, when I moved to Sorbonne 525 00:32:55,050 --> 00:32:57,784 to be in a very interdisciplinary environment. 526 00:32:57,792 --> 00:33:05,496 And at some point I, I, well, I stepped into the team. 527 00:33:05,504 --> 00:33:09,784 We have a very strong machine learning team here in the 528 00:33:09,792 --> 00:33:15,340 computer science department, and I discovered almost by chance 529 00:33:15,348 --> 00:33:19,121 that they were doing machine learning for physics. 530 00:33:19,129 --> 00:33:22,990 So I contacted them and said, OK, I'm a fluid mechanicist, I'm 531 00:33:22,990 --> 00:33:24,709 trying to move to machine learning. 532 00:33:24,717 --> 00:33:27,392 I know that you're doing machine learning and you're trying to 533 00:33:27,400 --> 00:33:30,480 move to fluid mechanics. Can we do something together? 534 00:33:30,488 --> 00:33:34,748 And so we applied together to an internal funding instrument of 535 00:33:34,756 --> 00:33:37,600 the university and we got funded. 536 00:33:37,608 --> 00:33:39,743 So we had this LearnFluidS team. 537 00:33:39,751 --> 00:33:46,075 So it was a little bit of money to have a post doc and a few 538 00:33:46,083 --> 00:33:50,754 interns and little things, but this allowed to connect each 539 00:33:50,762 --> 00:33:53,803 other. And then they said, Hey, we have 540 00:33:53,803 --> 00:33:57,776 this guy who is going to do machine learning methods for 541 00:33:57,784 --> 00:34:01,440 CFD. So we want to, but we have no 542 00:34:01,448 --> 00:34:02,840 databases. In machine learning. 543 00:34:02,840 --> 00:34:09,440 We, we have MNIST. We have, you know, many famous 544 00:34:09,440 --> 00:34:13,120 databases, but there are no good databases in CFD. 545 00:34:13,120 --> 00:34:16,538 We want to build 1, but actually we don't know exactly how to 546 00:34:16,538 --> 00:34:20,679 do the meshes because they were trying to use, you know, 547 00:34:20,679 --> 00:34:25,880 OpenFOAM, and we're using, you know, snappyHexMesh. 548 00:34:25,880 --> 00:34:28,760 Yeah. And of course the meshes are. 549 00:34:28,760 --> 00:34:30,516 Not very good close to the wall. 550 00:34:30,525 --> 00:34:33,795 And that's the point. So I said no, but you can also 551 00:34:33,803 --> 00:34:35,645 use the structured meshes and so on. 552 00:34:35,652 --> 00:34:40,000 And then, and then also we started to discuss a little bit 553 00:34:40,000 --> 00:34:43,016 about the criteria we had to use to evaluate the results. 554 00:34:43,024 --> 00:34:47,000 Not only the MSE we have discussed, but you know, and 555 00:34:47,007 --> 00:34:50,929 that's how well eventually they produce this AirfRANS. 556 00:34:50,938 --> 00:34:55,069 Of course, the well, the marriage is essentially the, 557 00:34:55,077 --> 00:34:57,690 the, the students. It's something interesting that 558 00:34:57,690 --> 00:35:01,144 this student had the background in physics the beginning, so and 559 00:35:01,144 --> 00:35:04,440 he moved to machine learning too, but he had a background in 560 00:35:04,440 --> 00:35:08,845 not in CFD but in physics. And so, yeah, so with this 561 00:35:08,853 --> 00:35:13,414 mixture of disciplines, we came up with this, with this 562 00:35:13,422 --> 00:35:15,946 database. And also they could test a lot 563 00:35:15,954 --> 00:35:18,742 of baselines that the way they do it in machine learning or 564 00:35:18,750 --> 00:35:21,858 they take 1 based and two baselines and they test all the 565 00:35:21,858 --> 00:35:25,640 baselines. And now we are trying to do more 566 00:35:25,640 --> 00:35:27,315 actually also for unsteady flows. 567 00:35:27,323 --> 00:35:32,538 So hopefully we will have a new database coming out which is on 568 00:35:32,538 --> 00:35:35,682 LES. So this times is not steady 569 00:35:35,690 --> 00:35:39,800 RANS, it's more LES and it's actually it snapshots because 570 00:35:39,808 --> 00:35:45,380 you know to to have time resolved predictions, yes. 571 00:35:45,388 --> 00:35:48,602 And. And the idea is the same. 572 00:35:48,610 --> 00:35:52,076 So they, they well enlighten me on machine learning 573 00:35:52,084 --> 00:35:54,260 architectures because they know better. 574 00:35:54,268 --> 00:36:00,015 They know, they know better. Also the, the, the sometimes 575 00:36:00,023 --> 00:36:06,055 the, you know, the, the, the, the pitfalls and the you can 576 00:36:06,063 --> 00:36:09,420 have in training these things because sometimes, you know, the 577 00:36:09,420 --> 00:36:15,196 training is not an easy. But on the other side, I say, 578 00:36:15,204 --> 00:36:19,092 OK, maybe you should look at this and that you should use 579 00:36:19,100 --> 00:36:20,960 this physical criteria and so on. 580 00:36:20,968 --> 00:36:24,130 And it's very, very instructive, I think, for both sides. 581 00:36:24,138 --> 00:36:26,400 So I'm very happy with this collaboration. 582 00:36:26,408 --> 00:36:34,976 And yeah, that's that's how we. I mean, how much have you 583 00:36:34,984 --> 00:36:39,432 struggled, though to get acceptance in that the from when 584 00:36:39,432 --> 00:36:42,560 you started to now? Have you how have you seen 585 00:36:42,568 --> 00:36:45,160 things progress? You know, the the argument of 586 00:36:45,168 --> 00:36:49,292 how well is the model just a fancy interpolation versus 587 00:36:49,300 --> 00:36:53,424 actually learning the physics? You know, how much is that is a 588 00:36:53,424 --> 00:36:57,635 a good thing that this is hard questions and how much of it is 589 00:36:57,635 --> 00:37:00,280 almost holding things back a little bit? 590 00:37:00,288 --> 00:37:02,300 Well, the. Hardest thing, Well, I started 591 00:37:02,308 --> 00:37:03,907 with the turbulence for this right? 592 00:37:03,915 --> 00:37:08,296 And it was hard beginning because, you know, well, 593 00:37:08,304 --> 00:37:11,784 turbulence models have a lot of knowledge. 594 00:37:11,792 --> 00:37:15,909 They are really incredible. They have this physics sense, 595 00:37:15,917 --> 00:37:23,280 which is, I don't know, but they are also very, you know, fond of 596 00:37:23,280 --> 00:37:25,854 their methodology. They, you know, the fact that 597 00:37:25,862 --> 00:37:30,146 you have do you have to do the things like in a certain way? 598 00:37:30,154 --> 00:37:34,480 And also the parameters has been, have been the model 599 00:37:34,488 --> 00:37:38,702 parameters been tuned making a lot of compromises and so on. 600 00:37:38,710 --> 00:37:42,060 So they don't like that you start playing with the 601 00:37:42,068 --> 00:37:45,720 parameters, playing with the terms and so on and so on. 602 00:37:45,728 --> 00:37:50,965 And for sure at the beginning the community of people who were 603 00:37:50,965 --> 00:37:54,980 who was playing with machine learning or, or calibration also 604 00:37:54,980 --> 00:37:58,936 calibrations of of turbulence model was not a community of 605 00:37:58,944 --> 00:38:02,420 turbulence modellers was a community of people coming like 606 00:38:02,420 --> 00:38:05,814 me from numerical schemes actually or from numerics in 607 00:38:05,822 --> 00:38:10,994 general. And so we didn't have the right 608 00:38:10,994 --> 00:38:15,650 codes for turbulence modelling, which is, you know, there is a 609 00:38:15,658 --> 00:38:18,380 lot of knowledge accumulated for for decades. 610 00:38:18,388 --> 00:38:23,320 And so many people were doing things that were not acceptable 611 00:38:23,320 --> 00:38:26,040 actually from a strict turbulence modeling point of 612 00:38:26,048 --> 00:38:28,302 view. So they were not using the right 613 00:38:28,302 --> 00:38:30,572 features as the input of the turbulent model. 614 00:38:30,580 --> 00:38:33,904 For instance, some people were using velocities at the 615 00:38:33,912 --> 00:38:38,435 beginning, which, you know, it's it's not Galilean invariant and 616 00:38:38,435 --> 00:38:42,824 so on. And and however, well, still 617 00:38:42,832 --> 00:38:46,540 this model looked interesting. So everybody was intrigued with 618 00:38:46,540 --> 00:38:49,692 the with the with the possibility of having it in this 619 00:38:49,692 --> 00:38:51,274 model. Because in the end, if you look 620 00:38:51,274 --> 00:38:54,641 at even if the to the traditional turbulence models, 621 00:38:54,649 --> 00:38:59,100 they remain data-driven. They yeah, data-driven with a 622 00:38:59,108 --> 00:39:01,416 human, you know, tuning the parameters. 623 00:39:01,424 --> 00:39:04,706 But they are data-driven actually because you are you are 624 00:39:04,706 --> 00:39:07,894 calibrating on, on some data sets which are the canonical 625 00:39:07,902 --> 00:39:10,615 flows. But OK, so at some point, well 626 00:39:10,623 --> 00:39:15,746 you had Chris Rumsey, which you know who, you know who decides 627 00:39:15,754 --> 00:39:22,005 to organize a meeting at it was in Virginia, it was at the was 628 00:39:22,013 --> 00:39:24,830 the name, the light, the lighthouse, right, the 629 00:39:24,838 --> 00:39:27,840 lighthouse. So there was this mythical 630 00:39:27,848 --> 00:39:32,252 meeting in 2022 where he said, OK, we are going to put around 631 00:39:32,260 --> 00:39:34,980 the table classical turbulence modeler and these guys, these, 632 00:39:34,988 --> 00:39:37,424 you know, these power venues with their machine learning 633 00:39:37,432 --> 00:39:41,948 stuff. And and it was really 634 00:39:41,956 --> 00:39:44,831 instructive because you had incredible guys there. 635 00:39:44,839 --> 00:39:49,176 So it was a in honor of the 60th birthday of Philippe Spalart, 636 00:39:49,176 --> 00:39:54,124 who is a great guy by the way. And you had, you know, Philippe, 637 00:39:54,124 --> 00:39:58,330 you had Paul Durbin, you had all these guys, you know, you 638 00:39:58,330 --> 00:40:01,521 know, plenty of things on turbulence modelling and they 639 00:40:01,521 --> 00:40:04,724 were explaining things. And on the other hand, we were 640 00:40:04,732 --> 00:40:07,280 trying to defend the idea of using machine learning. 641 00:40:07,288 --> 00:40:13,970 So we were a little bit, you know, like how, how do you say 642 00:40:13,978 --> 00:40:17,035 you attack? Attack. 643 00:40:17,043 --> 00:40:20,652 But it was instructive because I think thanks to that, we 644 00:40:20,660 --> 00:40:24,052 progressed a lot because we, we started with the idea, OK, it's 645 00:40:24,052 --> 00:40:27,400 nice. We just fine tune a model for a 646 00:40:27,400 --> 00:40:30,180 very narrow set of flows. We get better results. 647 00:40:30,188 --> 00:40:34,384 We are happy with that. And now we are moving more and 648 00:40:34,392 --> 00:40:38,464 more toward unifying models. And maybe, maybe, maybe I, I 649 00:40:38,472 --> 00:40:40,670 maybe show ambition, but never know. 650 00:40:40,678 --> 00:40:45,715 Maybe in 10 years more we could have a foundation model that's a 651 00:40:45,715 --> 00:40:49,464 dream and turbulence model which could, you know, actually 652 00:40:49,464 --> 00:40:54,630 realize the the the dream of turbulence model of a universal 653 00:40:54,630 --> 00:40:59,055 turbulence model. I'm not sure we will get there, 654 00:40:59,055 --> 00:41:04,200 but you know it is. Interesting though, because in 655 00:41:04,208 --> 00:41:10,648 some ways turbulence modelling was going out of fashion. 656 00:41:10,656 --> 00:41:15,296 And I know that it used to be the joke, didn't it? 657 00:41:15,304 --> 00:41:17,940 I think that for a European project, if you said you're 658 00:41:17,948 --> 00:41:20,468 going to work on turbulence modelling, it was almost like 659 00:41:20,476 --> 00:41:24,274 guaranteed to be rejected because it was seen as a solved 660 00:41:24,274 --> 00:41:29,918 problem or you know what's new? And but there were the the irony 661 00:41:29,918 --> 00:41:34,737 is is the industry still uses mainly RANS and are stuck using 662 00:41:34,737 --> 00:41:38,880 methods from the 80s or or early 90s seems like with 663 00:41:38,880 --> 00:41:42,712 machine learning. Then there was this spike of of 664 00:41:42,712 --> 00:41:46,092 of potential again. But I don't know about you, but 665 00:41:46,092 --> 00:41:49,756 I almost felt that. Maybe some of the? 666 00:41:49,764 --> 00:41:53,320 Use of machine learning for turbulence modeling was a little 667 00:41:53,320 --> 00:41:58,335 bit early and the expectation was so big that when it didn't 668 00:41:58,343 --> 00:42:02,216 meet that expectation, it sort of then dropped down again. 669 00:42:02,224 --> 00:42:08,263 And, and I, I, but I agree with you that in some ways with the 670 00:42:08,271 --> 00:42:12,960 idea that we want to build also foundation models from a 671 00:42:12,968 --> 00:42:16,368 surrogate modelling side, having to do everything with LES or 672 00:42:16,376 --> 00:42:18,595 whatever, it's just so computationally expensive. 673 00:42:18,603 --> 00:42:22,915 Ironically, if you could use machine learning to come up with 674 00:42:22,915 --> 00:42:26,750 the ultimate turbulence model, you would actually make it much 675 00:42:26,750 --> 00:42:30,708 more affordable to, to, to actually run the simulations to 676 00:42:30,708 --> 00:42:33,080 achieve. So I feel like now there is 677 00:42:33,080 --> 00:42:38,654 maybe a little bit more economic or relevant again of the 678 00:42:38,654 --> 00:42:45,233 turbulence modelling because it would save so much on the date 679 00:42:45,233 --> 00:42:51,000 generation side. But yeah, it it does seem to be, 680 00:42:51,600 --> 00:42:55,800 I don't know if you've noticed, but I saw in CFD anyway that 681 00:42:55,800 --> 00:42:58,432 initially everybody was focused on machine learning for 682 00:42:58,432 --> 00:43:03,894 turbulence models, where now it seems to be far more about 683 00:43:03,894 --> 00:43:06,684 machine learning for surrogate models. 684 00:43:06,692 --> 00:43:13,320 And that seems to be much more focused on and you don't hear as 685 00:43:13,320 --> 00:43:16,740 much on maybe the turbulence modelling in such a strong way. 686 00:43:16,740 --> 00:43:18,360 Would would, would you tend to 687 00:43:18,360 --> 00:43:21,400 agree with that, that the community sort of shifted a 688 00:43:21,400 --> 00:43:25,960 little bit? Some communities, yes, I think 689 00:43:25,960 --> 00:43:31,840 in the, in the well in, in complex engineering applications 690 00:43:32,000 --> 00:43:38,600 like well car industry or even in solid mechanics for instance, 691 00:43:38,600 --> 00:43:40,760 they are shifting in that direction. 692 00:43:41,760 --> 00:43:44,280 I think that in aerospace they are still interesting to 693 00:43:44,280 --> 00:43:47,360 turbulence models, especially for complex application like 694 00:43:47,360 --> 00:43:52,800 hypersonics or transition models, all that. 695 00:43:54,320 --> 00:44:01,040 That's because, well, even if you can, well what you would 696 00:44:01,040 --> 00:44:06,520 like to have in in for instance in aerospace where we have what 697 00:44:06,520 --> 00:44:11,200 you would like to have is a high fidelity model, so like LES 698 00:44:11,200 --> 00:44:13,960 quality and to perform your optimization with that. 699 00:44:13,960 --> 00:44:16,120 Why? Because we know that RANS has 700 00:44:16,120 --> 00:44:20,960 flaws, because we want to explore extreme, extreme 701 00:44:20,960 --> 00:44:25,712 operating conditions, because we are moving away from known 702 00:44:25,712 --> 00:44:29,480 paths like we are changing the fuels, we are changing the 703 00:44:29,480 --> 00:44:34,720 architectures and so on. So you cannot just design, I 704 00:44:34,720 --> 00:44:38,520 don't know, a new propeller or a new wing based on epsilon 705 00:44:38,520 --> 00:44:41,960 modifications of something you know and for which you know that 706 00:44:41,960 --> 00:44:44,680 the turbulence model is going to be wrong, but you know more or 707 00:44:44,680 --> 00:44:48,000 less how it's going to fail and how you should correct it. 708 00:44:48,280 --> 00:44:52,000 OK. And the problem is that even if 709 00:44:52,000 --> 00:44:55,760 right now with the, the, the, the porting to GPUs and so on, 710 00:44:56,760 --> 00:45:00,720 high fidelity simulation are becoming more affordable, still 711 00:45:00,720 --> 00:45:04,200 that's one simulation. If you want to perform an 712 00:45:04,200 --> 00:45:08,680 optimization, you need thousands of LES. 713 00:45:08,840 --> 00:45:12,791 And even if you can run a complex wall-modelled LES on on 714 00:45:12,791 --> 00:45:18,960 a GPU, what you know, chatting, you know, well, you know this, 715 00:45:20,440 --> 00:45:23,320 but still it's 1. So if you want to perform, I 716 00:45:23,320 --> 00:45:28,160 don't know, 2000 simulations, you need 2000 GPUs, which is 717 00:45:28,160 --> 00:45:31,240 very costly. I don't know how many companies 718 00:45:31,240 --> 00:45:35,920 can, you know, afford 2000 GPUs, But there are and also, if you 719 00:45:35,920 --> 00:45:38,920 have, you know, few GPUs, well, you have to wait for many, many 720 00:45:38,920 --> 00:45:43,720 days and the design cycles to be shorter for for economical 721 00:45:43,720 --> 00:45:46,600 reasons and also for environmental reasons. 722 00:45:46,600 --> 00:45:51,080 I mean, OK, so for that you cannot rely on LES alone. 723 00:45:51,080 --> 00:45:55,480 So what you can try to do is to distill the knowledge of this 724 00:45:55,480 --> 00:45:58,440 high fidelity method into lower order models. 725 00:45:59,000 --> 00:46:03,040 So the how can you do it? That's the way I tried to do it. 726 00:46:03,400 --> 00:46:08,360 One possibility is to distill this into augmented RANS models. 727 00:46:08,480 --> 00:46:12,160 And then once you have augmented your RANS model, you use that 728 00:46:12,160 --> 00:46:15,360 one to perform the optimization cycle, which is much more 729 00:46:15,360 --> 00:46:20,600 affordable, provided that your model generalizes well enough, 730 00:46:20,600 --> 00:46:26,120 at least on your design space. Because if the model fails as 731 00:46:26,120 --> 00:46:28,960 you move a little bit far from the baseline, well, that's of 732 00:46:28,960 --> 00:46:32,680 course that's useless. The other point you can do is to 733 00:46:32,880 --> 00:46:35,840 surrogate models. OK, but the problem with 734 00:46:35,840 --> 00:46:40,160 surrogate model is that you will never have enough LES to train 735 00:46:40,160 --> 00:46:42,128 a surrogate model on it on LES. 736 00:46:42,136 --> 00:46:44,263 So you are trying to do that right? 737 00:46:44,271 --> 00:46:48,140 You are trying to produce high fidelity databases but they are 738 00:46:48,140 --> 00:46:52,820 still relatively limited and I don't know how many of them will 739 00:46:52,820 --> 00:46:55,640 be possible to produce especially for very complex very 740 00:46:55,640 --> 00:46:59,712 high Reynolds number flows. And if we have so many data at 741 00:46:59,720 --> 00:47:03,340 some point, what's the point of having surrogates if we can run 742 00:47:03,340 --> 00:47:06,747 thousands and 10s of thousands of LES, what's the point of 743 00:47:06,755 --> 00:47:08,856 having a surrogate? So I don't know. 744 00:47:08,864 --> 00:47:14,006 So maybe what's what we are trying to do is to have multi 745 00:47:14,014 --> 00:47:17,450 fidelity models. So where actually you train the 746 00:47:17,450 --> 00:47:21,918 model using data of different origins provided that you have a 747 00:47:21,918 --> 00:47:27,282 clear hierarchy and you can actually try to to train using 748 00:47:27,290 --> 00:47:30,818 plenty of low fidelity data like RANS. 749 00:47:30,826 --> 00:47:35,138 So that's where having a RANS, which is not too bad remains 750 00:47:35,146 --> 00:47:38,718 useful. And then you try to learn the 751 00:47:38,726 --> 00:47:43,070 gap between the RANS and the high fidelity and then you use 752 00:47:43,078 --> 00:47:46,634 your surrogate. And this is something which is 753 00:47:46,642 --> 00:47:51,932 not yet so used in the surrogate modelling community because they 754 00:47:51,932 --> 00:47:57,720 either use plenty of RANS data or well, whoever some databases, 755 00:47:57,720 --> 00:48:01,292 high fidelity database like the one you you have contributed 756 00:48:01,292 --> 00:48:04,680 to produce. But still in a database you have 757 00:48:04,680 --> 00:48:08,692 what, 300 cases for a family of cases, But it's not enough 758 00:48:08,692 --> 00:48:12,588 to, you know, to have something which can be reused 759 00:48:12,588 --> 00:48:15,996 for anything. So this means that every time 760 00:48:16,004 --> 00:48:20,588 you change, you need to run 300 and yes, or, or one model that 761 00:48:20,596 --> 00:48:22,907 yes or whatever. And that's costly. 762 00:48:22,915 --> 00:48:27,800 And yeah, So no, I, I. That's kind of why I think there 763 00:48:27,800 --> 00:48:32,600 is a connection between the two for sure that the the the 764 00:48:32,608 --> 00:48:38,250 underlying CFD is still the key and making that more affordable. 765 00:48:38,250 --> 00:48:39,600 Yeah, you're right. 766 00:48:39,600 --> 00:48:42,880 There are start-ups like like Volcano and others who are 767 00:48:42,880 --> 00:48:45,520 trying to come up with very computationally efficient codes. 768 00:48:46,080 --> 00:48:51,160 But that is kind of the whole point of turbulence modelling in 769 00:48:51,160 --> 00:48:54,560 some ways is to, is to try to come up with a way of modelling 770 00:48:54,560 --> 00:48:57,600 it in a, in a, in a, in a clever way. 771 00:48:57,720 --> 00:49:02,000 So, yeah, if it, if it can be done, then that would be then 772 00:49:02,000 --> 00:49:05,600 that would be good. I, I did want to ask you on the, 773 00:49:05,600 --> 00:49:09,080 I mean, you're an editor in chief of Computers & Fluids. 774 00:49:09,080 --> 00:49:13,371 You're extremely active in the, you know, academic world when 775 00:49:13,371 --> 00:49:18,040 it comes to publishing. You know, how how are things 776 00:49:18,040 --> 00:49:20,920 changing with the rise of like ML based papers? 777 00:49:21,320 --> 00:49:24,800 You know, what standards should they be having? 778 00:49:24,800 --> 00:49:26,920 You know, you would previously we'd always say, I want to see a 779 00:49:26,920 --> 00:49:28,640 mesh requirement study. I want to see proof of 780 00:49:28,640 --> 00:49:32,400 convergence. I want to see, you know what, 781 00:49:33,520 --> 00:49:36,280 what standards are you wanting and seeing? 782 00:49:36,280 --> 00:49:38,840 And and yeah, I'd be interested to hear your thoughts as a 783 00:49:38,840 --> 00:49:40,960 journal editor. Yeah, what we are. 784 00:49:40,960 --> 00:49:46,120 Doing right now in Computers & Fluids is to ask to motivate 785 00:49:46,120 --> 00:49:51,320 very well why you do need machine learning and why this is 786 00:49:51,320 --> 00:49:55,560 improving over, I would say, standard approaches. 787 00:49:56,680 --> 00:50:00,280 To give an example, many people do OK, we were talking about 788 00:50:00,280 --> 00:50:04,680 surrogates and they say, OK, I want to design a new airfoil. 789 00:50:05,120 --> 00:50:10,160 So new airfoil and for that I generated a database of 10,000 790 00:50:10,160 --> 00:50:14,040 RANS simulations and then I trained the surrogate and then 791 00:50:14,240 --> 00:50:21,520 here goes my my, my, my optimal profile optimized using the 792 00:50:21,520 --> 00:50:24,120 surrogate. But if you have the 793 00:50:24,160 --> 00:50:29,680 computational power to run 10,000 RANS, you can do brute 794 00:50:29,680 --> 00:50:32,880 force optimization. You don't need actually the 795 00:50:32,880 --> 00:50:36,640 surrogates, right. So what we are trying to ask is 796 00:50:36,640 --> 00:50:41,080 to show clearly that you are learning something more than. 797 00:50:41,720 --> 00:50:45,320 So your method is, is bringing something more either in term of 798 00:50:46,320 --> 00:50:49,240 total computational time, including the time required to 799 00:50:49,240 --> 00:50:55,560 generate the database and in terms of improved design or 800 00:50:55,560 --> 00:51:00,600 showing that you, you know, you are making up for some failure 801 00:51:00,600 --> 00:51:04,720 of the classical model because otherwise showing OK, I train 802 00:51:05,120 --> 00:51:08,200 surrogate, it works well. Here's the optimum. 803 00:51:08,240 --> 00:51:12,200 It's, you know, it's, it's not I, I believe it, it's, it's, 804 00:51:12,200 --> 00:51:14,920 it's better. But you know, you are not 805 00:51:15,840 --> 00:51:19,840 telling the community what's new with respect to things we, we 806 00:51:20,040 --> 00:51:24,080 already are able to do. For instance, in some cases we 807 00:51:24,080 --> 00:51:28,480 have tested some baselines on airfoils, you can generate an 808 00:51:28,480 --> 00:51:31,960 airfoil solution of full-field RANS in 5 minutes. 809 00:51:31,960 --> 00:51:34,240 You know using FreeFEM or any any code. 810 00:51:34,400 --> 00:51:42,800 OK to train some surrogates you need 20 or 25 or 30 hours of of 811 00:51:42,800 --> 00:51:46,840 GPU time. So in 20 hours I have largely 812 00:51:46,840 --> 00:51:49,800 finished my optimization with my old RANS code. 813 00:51:50,080 --> 00:51:54,760 So the point is to show that either you go to more complex 814 00:51:54,760 --> 00:51:56,760 models and you can do that fastly. 815 00:51:57,120 --> 00:52:00,600 So you have, you know, each RANS simulation would take, I don't 816 00:52:00,600 --> 00:52:05,920 know, 100 hours and thanks to the surrogates, I'm speeding up 817 00:52:05,920 --> 00:52:09,840 and I can have it in, I don't know, 20 hours or 24 hours or 818 00:52:09,840 --> 00:52:12,360 it's not new. So what we are asking is that 819 00:52:12,360 --> 00:52:18,080 OK, not only show that your model has been trained well and 820 00:52:18,080 --> 00:52:22,371 it performs better, but also try to compare to classical 821 00:52:22,371 --> 00:52:26,783 methods, show what's wrong with the classical method, what's 822 00:52:26,783 --> 00:52:30,440 what what your new machine learning method is bringing. 823 00:52:30,880 --> 00:52:34,560 And also show eventually where your machine learning method is 824 00:52:34,560 --> 00:52:37,720 going to fail because that's also important, not only 825 00:52:37,720 --> 00:52:41,880 showing, you know, positive results, but also negative 826 00:52:41,880 --> 00:52:43,480 results. And that's something the 827 00:52:43,480 --> 00:52:46,360 community doesn't like to do too much because it's easier to 828 00:52:46,360 --> 00:52:48,320 publish when you have good results. 829 00:52:48,320 --> 00:52:51,120 You know that's. That's a very good point 830 00:52:51,120 --> 00:52:53,520 actually. I, I don't, I think if you look 831 00:52:53,520 --> 00:53:00,284 at like NeurIPS or those papers conferences, they mandate like 832 00:53:00,284 --> 00:53:05,000 a limitations section where you have to put down all the bits 833 00:53:05,000 --> 00:53:08,320 where you didn't do it. And unless I'm mistaken, that 834 00:53:09,080 --> 00:53:12,280 that precedent of very clearly pointing out where the 835 00:53:12,280 --> 00:53:15,640 limitation is not something that's traditionally done right 836 00:53:15,720 --> 00:53:21,680 in not in such a clear before the conclusions limitations. 837 00:53:21,680 --> 00:53:25,560 And yeah, I've often felt that there needs to be more 838 00:53:25,560 --> 00:53:28,920 transparency also from just the CFD side where things are bad, 839 00:53:29,400 --> 00:53:33,800 you know, show me examples of machine learning, for example, 840 00:53:33,800 --> 00:53:37,800 where the where if I pick a certain split or, or I pick a 841 00:53:37,800 --> 00:53:39,576 thing, I get really bad results. 842 00:53:39,584 --> 00:53:43,160 It's actually quite important. Or else the conclusion is that 843 00:53:43,168 --> 00:53:48,116 these models are amazing and standard CFD's gone rather than 844 00:53:48,116 --> 00:53:52,292 pointing out they're positives and and where they you failed. 845 00:53:52,300 --> 00:53:56,580 The classic one I guess is in distribution and out of 846 00:53:56,588 --> 00:54:00,166 distribution where like yeah. Exactly. 847 00:54:00,174 --> 00:54:04,663 And the other point is to evaluate the models is as also 848 00:54:04,671 --> 00:54:08,855 discussion. We, we have had as aware compare 849 00:54:08,855 --> 00:54:14,180 the model on CFD criteria, not only on machine learning 850 00:54:14,188 --> 00:54:17,740 criteria, because OK, mean squared errors are the standard 851 00:54:17,740 --> 00:54:20,695 in machine learning. They are useful, whatever, but 852 00:54:20,703 --> 00:54:25,861 they can be also, you know, misleading, because in 853 00:54:25,869 --> 00:54:30,160 particular in external aerodynamics, where most of the 854 00:54:30,160 --> 00:54:34,758 flow is uniform, basically you are training a neural network to 855 00:54:34,758 --> 00:54:37,560 capture a constant function, and the regions where 856 00:54:37,560 --> 00:54:41,271 something's going on which are close to the wall are very tiny. 857 00:54:41,271 --> 00:54:43,840 So if you are wrong in that 858 00:54:43,840 --> 00:54:48,320 region comparatively to the whole with all the points you 859 00:54:48,320 --> 00:54:52,680 have, maybe it's not enough to have a, a clear signal on the 860 00:54:52,680 --> 00:54:55,120 MSE. But you see it if you plot 861 00:54:55,120 --> 00:54:58,640 velocity profiles, if you plot pressure distribution in 862 00:54:58,640 --> 00:55:02,760 general, they are very bumpy. If you plot skin friction, skin 863 00:55:02,760 --> 00:55:05,400 friction is terrible because in that case you are not only 864 00:55:05,400 --> 00:55:08,800 reconstructing the flow field, the the the velocity field, you 865 00:55:08,800 --> 00:55:11,480 are taking the derivatives. And it's well known that 866 00:55:11,480 --> 00:55:14,280 approximating the derivative of a function is more difficult 867 00:55:14,280 --> 00:55:16,440 than approximating the function itself, right? 868 00:55:17,000 --> 00:55:21,680 And so in the machine learning community is not used to this 869 00:55:21,680 --> 00:55:24,440 criteria, which are the standard criteria in CFD. 870 00:55:24,800 --> 00:55:30,520 So again, I I think that if you pretend to, if you, if you say 871 00:55:30,520 --> 00:55:33,880 that you are going to replace standard CFD approaches with 872 00:55:33,880 --> 00:55:36,971 machine learning, then you have to show that machine learning 873 00:55:36,971 --> 00:55:39,786 is able to provide the same thing that the CFD group is 874 00:55:39,786 --> 00:55:42,520 able to provide and funny. Enough. 875 00:55:42,520 --> 00:55:46,320 I was having this discussion on a, on a, on a paper we're 876 00:55:46,320 --> 00:55:51,640 putting out where it's also the case, isn't it, that in machine 877 00:55:51,640 --> 00:55:58,040 learning, because you have, let's say, 300 or 100 or 50 test 878 00:55:58,600 --> 00:56:01,565 cases, you'll average over them. 879 00:56:01,573 --> 00:56:09,370 And as your number where that has the same risk of some are 880 00:56:09,378 --> 00:56:12,085 really bad, some are really good. 881 00:56:12,093 --> 00:56:13,515 And then you have lots in the middle. 882 00:56:13,523 --> 00:56:17,232 Or is it you have lots, you know, how, how bad is bad, how 883 00:56:17,240 --> 00:56:20,560 good is good? And, and this isn't fully 884 00:56:20,568 --> 00:56:23,578 explained just by a single number. 885 00:56:23,586 --> 00:56:29,984 So this, as you say, almost doing a deeper analysis and 886 00:56:29,992 --> 00:56:35,758 showing I, I guess that what someone described to me is they 887 00:56:35,758 --> 00:56:40,864 to, to the early point of like hype because of the way the 888 00:56:40,872 --> 00:56:45,742 results are presented today, where an R-squared value or even 889 00:56:45,742 --> 00:56:49,520 L2 gives such low errors, it looks like they're perfect. 890 00:56:49,528 --> 00:56:52,990 But then when they try and practice and they don't see as 891 00:56:52,998 --> 00:56:56,525 good, it almost creates A disappointment, which if the 892 00:56:56,533 --> 00:57:00,040 paper had shown it, they probably would not be 893 00:57:00,048 --> 00:57:03,086 disappointed because it would just be meeting the expectation 894 00:57:03,086 --> 00:57:05,621 where yes. So I would agree with you, part 895 00:57:05,621 --> 00:57:09,514 of that is down to the way that the results have been presented. 896 00:57:09,514 --> 00:57:09,840 Yeah. 897 00:57:10,480 --> 00:57:13,240 There is quite a lot of overselling, but that's because 898 00:57:13,240 --> 00:57:16,720 of the economical model also to continue, you know, and, and 899 00:57:16,720 --> 00:57:20,480 the, and the publication system, which is a bit too, you know, 900 00:57:21,240 --> 00:57:23,880 especially in, in some countries, that's not the case 901 00:57:23,880 --> 00:57:26,040 in France, but in some countries, there's this huge 902 00:57:26,040 --> 00:57:29,080 pressure for publication for being the 1st for, you know, 903 00:57:29,520 --> 00:57:32,840 this is the store. And I think this, this, this 904 00:57:33,800 --> 00:57:38,320 rush to publication publishing more and more in higher impact 905 00:57:38,320 --> 00:57:44,280 in journals and so on, in the end is increasing in the, the 906 00:57:44,280 --> 00:57:50,360 signal to noise ratio at, at such a point that people don't 907 00:57:50,360 --> 00:57:53,640 even read the papers. They just read reviews of paper 908 00:57:53,640 --> 00:57:56,240 which are automatically generated by AIs. 909 00:57:56,520 --> 00:57:59,640 So there is something fundamentally wrong in this 910 00:57:59,680 --> 00:58:03,240 because, you know, which should remain at a reasonable level 911 00:58:03,240 --> 00:58:07,560 where humans are talking to humans because, you know, we're 912 00:58:07,560 --> 00:58:10,160 trying to produce human knowledge in the end. 913 00:58:10,240 --> 00:58:14,640 Because what's the point if all the AI is told to each other, 914 00:58:14,920 --> 00:58:17,680 hey, what are we going to do as you? 915 00:58:18,600 --> 00:58:21,120 I'm going to look at the beach. 916 00:58:21,128 --> 00:58:26,690 What the beach is not really not bad, but stay on the beach all 917 00:58:26,698 --> 00:58:30,150 the year. You have to find something else 918 00:58:30,150 --> 00:58:34,175 to do. So, yeah, no, I think that we 919 00:58:34,183 --> 00:58:39,172 should maybe slow down a little bit and do what well people used 920 00:58:39,172 --> 00:58:44,640 to do some years ago, which means publish less, but publish 921 00:58:44,640 --> 00:58:49,920 more thoughtful papers and wait a little bit before publishing. 922 00:58:49,920 --> 00:58:52,240 And when you publish, publish 923 00:58:52,240 --> 00:58:54,600 something which is more complete, you have looked to all 924 00:58:54,600 --> 00:59:00,480 the consequences and you know, thought to all the possible yes. 925 00:59:01,040 --> 00:59:04,480 Do you know what you mean? This is the, the good and the 926 00:59:04,480 --> 00:59:07,840 bad thing of things like arXiv where there's almost a sense of 927 00:59:08,680 --> 00:59:11,680 let's get it out there. We want people to see it. 928 00:59:11,680 --> 00:59:14,560 It's a preprint. The preprint supposed to be 929 00:59:14,560 --> 00:59:17,360 taken as well. You know, this hasn't been 930 00:59:17,360 --> 00:59:21,000 reviewed yet. But unfortunately people just 931 00:59:21,000 --> 00:59:25,560 take it as the main thing. And because things are moving so 932 00:59:25,560 --> 00:59:28,920 fast, the, the fact that it hasn't been reviewed is 933 00:59:28,920 --> 00:59:32,840 sometimes almost ignored, you know, by, by, by people in a 934 00:59:32,840 --> 00:59:36,640 way. So it's, I agree with you there, 935 00:59:37,960 --> 00:59:42,160 the pace necessitates sometimes to put it out, but at the same 936 00:59:42,160 --> 00:59:45,600 time that feeling in your back at the head of going, well, I 937 00:59:45,600 --> 00:59:48,440 could really do with another month or two to really deeply 938 00:59:48,440 --> 00:59:51,680 investigate this. Is. 939 00:59:52,080 --> 00:59:55,600 Balanced against the desire just to get something out, it's. 940 00:59:56,320 --> 00:59:58,080 It's a tricky 1 and. On the other. 941 00:59:58,080 --> 01:00:02,560 Hand the editorial The standard editorial system is being 942 01:00:02,640 --> 01:00:06,000 flooded by papers. So it's becoming more and more 943 01:00:06,000 --> 01:00:10,000 difficult to keep quality in the review process. 944 01:00:11,920 --> 01:00:17,160 You know when when you are flooded by 10s of papers every 945 01:00:17,160 --> 01:00:20,440 day, OK. And you have to process them and 946 01:00:20,440 --> 01:00:22,520 you have to send them to reviewers. 947 01:00:22,520 --> 01:00:25,840 You have to find reviewers who are also flooded by other 948 01:00:25,840 --> 01:00:29,400 journals because there are many journals, OK, many, many 949 01:00:29,400 --> 01:00:32,960 journals and more journals are, are being created and so on. 950 01:00:33,440 --> 01:00:39,280 So at some point it's just, you know, just an escalation and 951 01:00:39,280 --> 01:00:42,000 it's very difficult to control the quality of the papers and 952 01:00:42,320 --> 01:00:47,120 what in, in fluid mechanics, things are becoming difficult, 953 01:00:47,120 --> 01:00:50,640 but they are still under control to some extent because it's a 954 01:00:50,800 --> 01:00:53,960 small community. But you have seen the examples 955 01:00:53,960 --> 01:00:58,280 of these big conferences like NeurIPS where they have 10s of 956 01:00:58,280 --> 01:01:03,120 thousands of papers, where most of the papers are written by AI 957 01:01:03,120 --> 01:01:08,920 and reviewed by AI. And so what's, what's the 958 01:01:08,920 --> 01:01:11,840 meaning of the review process in that case? 959 01:01:11,840 --> 01:01:15,240 You don't review it at all. You just put it for free on the 960 01:01:15,240 --> 01:01:17,840 website and who wants to read it? 961 01:01:18,840 --> 01:01:22,880 So I think that the, the, the rush to having more and more and 962 01:01:22,880 --> 01:01:27,640 more papers, which is also due to the, you know, to the 963 01:01:27,640 --> 01:01:31,480 publishers is in some sense killing system. 964 01:01:31,480 --> 01:01:36,600 Because if the review level falls below a given pressure, at 965 01:01:36,600 --> 01:01:39,644 some point, there is no point in going through a standard 966 01:01:39,644 --> 01:01:43,304 review process because there is no added values in going 967 01:01:43,304 --> 01:01:47,440 through that review. So you just produce your your 968 01:01:47,440 --> 01:01:50,800 paper, you put it somewhere and the people look at it if they 969 01:01:50,800 --> 01:01:54,400 want and that's it. And if you are in the hype, you 970 01:01:54,400 --> 01:01:58,080 are in, let's say, a famous team, very visible and so on, 971 01:01:58,080 --> 01:02:01,960 people will have followers who will read your paper because 972 01:02:01,960 --> 01:02:04,520 it's you. So we are shifting from 973 01:02:04,520 --> 01:02:08,320 scientists to influencers. And that's a bit, you know, 974 01:02:08,320 --> 01:02:10,120 scary for me. Yeah. 975 01:02:10,560 --> 01:02:15,760 Yeah, no, no the. So how does this, you know, 976 01:02:15,760 --> 01:02:19,440 translate and what it what's what are your trying to achieve 977 01:02:19,440 --> 01:02:23,743 with with, you know, the new position at Sorbonne AI centre? 978 01:02:23,743 --> 01:02:24,840 Could you maybe tell a little 979 01:02:24,840 --> 01:02:28,920 bit more about your place and and kind of your vision to how 980 01:02:28,920 --> 01:02:33,200 to do, I guess, AI for science and AI for engineering and and 981 01:02:33,200 --> 01:02:37,080 and maybe take some of those rigorous things that you've 982 01:02:37,640 --> 01:02:41,320 taken from the fluid mechanics world in into this side? 983 01:02:41,320 --> 01:02:43,120 That'd be. I love your shiny new office So 984 01:02:43,120 --> 01:02:46,760 I wanted to hear more. Yeah. 985 01:02:47,280 --> 01:02:50,440 Well, actually, well, in Sorbonne we have this centre, 986 01:02:50,440 --> 01:02:53,080 SCAI Sorbonne. The beginning, the name was 987 01:02:53,080 --> 01:02:56,320 Sorbonne Centre for Artificial Intelligence, which was created 988 01:02:56,320 --> 01:03:04,360 seven years ago at the beginning of the AI hype, because we had 989 01:03:04,360 --> 01:03:08,160 the strong teams in computer science and also in applied 990 01:03:08,160 --> 01:03:11,160 mathematics working on that. So basically at the beginning, 991 01:03:11,240 --> 01:03:16,460 SCAI has been created by applied mathematicians and and 992 01:03:16,460 --> 01:03:20,001 computer scientists alongside with some people from 993 01:03:20,001 --> 01:03:23,104 humanities and from other sciences, for instance of 994 01:03:23,112 --> 01:03:25,264 computational biology, this kind of thing. 995 01:03:25,272 --> 01:03:29,712 And well, it has helped a lot in in making. 996 01:03:29,720 --> 01:03:33,543 So in making AI, it's a one visible. 997 01:03:33,551 --> 01:03:37,725 They have generated a lot of activities we have. 998 01:03:37,733 --> 01:03:41,646 So at at some point we have been labeled by the French government 999 01:03:41,646 --> 01:03:44,996 as clusters. So now the C in SCAI stands for 1000 01:03:44,996 --> 01:03:48,852 cluster. And these clusters are sort of 1001 01:03:48,860 --> 01:03:53,465 centres of excellence in France. There are nine of them which are 1002 01:03:53,465 --> 01:04:00,823 supposed to foster research in AI training and transfer to to 1003 01:04:00,831 --> 01:04:05,940 companies. And well, every cluster has its 1004 01:04:05,940 --> 01:04:08,530 own specificities. But when? 1005 01:04:08,538 --> 01:04:14,242 Well, based on my own experience and based on the very large 1006 01:04:14,250 --> 01:04:17,795 scientific community at Sorbonne, I said, well, right 1007 01:04:17,803 --> 01:04:23,120 now everybody is looking at AI for science because science is 1008 01:04:23,128 --> 01:04:25,960 actually going to be deeply modified. 1009 01:04:25,968 --> 01:04:29,735 The way of doing science is going to be deeply modified by 1010 01:04:29,743 --> 01:04:34,624 the introduction of these AI approaches, which is not the 1011 01:04:34,632 --> 01:04:37,770 machine learning approaches for, you know, improving predictions 1012 01:04:37,770 --> 01:04:43,464 or speeding up predictions, but also reasoning tools and, you 1013 01:04:43,472 --> 01:04:50,647 know, models for, you know, for, for automating labs, 1014 01:04:50,655 --> 01:04:55,000 automating labs, AI agents and, and all that. 1015 01:04:55,008 --> 01:04:58,546 And I said, well, the, the difficulty with that is that 1016 01:04:58,554 --> 01:05:01,680 when you apply this to science, you have several problems. 1017 01:05:01,688 --> 01:05:05,272 First of all, you are not dealing with random data. 1018 01:05:05,280 --> 01:05:11,904 It's not images, it's not video. It's something which has a deep 1019 01:05:11,904 --> 01:05:16,200 physical meaning, which he's known only to experts from the 1020 01:05:16,208 --> 01:05:18,970 discipline. So you can train a surrogate or 1021 01:05:18,970 --> 01:05:22,020 whatever, or a foundation model using, I don't know, chemistry 1022 01:05:22,028 --> 01:05:24,362 data. But if you don't have a 1023 01:05:24,370 --> 01:05:28,364 chemist by you to explain, if what's coming out of the model 1024 01:05:28,372 --> 01:05:32,870 is you know, makes sense from a chemical point of view or is 1025 01:05:32,878 --> 01:05:35,020 just an hallucination, you don't know. 1026 01:05:35,028 --> 01:05:39,565 So that's where it's extremely important to connect people from 1027 01:05:39,565 --> 01:05:43,005 computer science and the disciplines and also people from 1028 01:05:43,005 --> 01:05:46,304 mathematics because they can help with the mathematical 1029 01:05:46,312 --> 01:05:49,271 foundations. What I dream about of is the 1030 01:05:49,279 --> 01:05:52,312 equivalence, you know, Lax equivalence theorem for machine 1031 01:05:52,312 --> 01:05:54,416 learning. That would be the grail, you 1032 01:05:54,424 --> 01:05:57,181 know, something that ensures that you, if you put more data 1033 01:05:57,189 --> 01:06:00,207 in your model, at some point you are going to converge to 1034 01:06:00,215 --> 01:06:02,778 something. But you know, that would be 1035 01:06:02,786 --> 01:06:05,296 great. But so the idea would be to have 1036 01:06:05,296 --> 01:06:06,623 all these people discussing together. 1037 01:06:06,631 --> 01:06:10,216 And then I realized that some problems you can have in one 1038 01:06:10,224 --> 01:06:13,360 discipline, for instance, in high energy physics, you have 1039 01:06:13,368 --> 01:06:17,436 big data. This data are structured, they 1040 01:06:17,444 --> 01:06:21,906 encode causalities because the particle I'm capturing here at 1041 01:06:21,914 --> 01:06:26,440 this moment comes from a jet which has originated there and 1042 01:06:26,448 --> 01:06:30,100 the particle have undergone some decay process and so on. 1043 01:06:30,108 --> 01:06:32,664 And this, they call this actually jets of particles. 1044 01:06:32,672 --> 01:06:36,039 And they say, oh, this is the same you can have in a turbulent 1045 01:06:36,039 --> 01:06:39,620 jet because the jet is originating from, you know, as 1046 01:06:39,628 --> 01:06:44,888 lice or something as law. And then you have causalities of 1047 01:06:44,888 --> 01:06:47,010 turbulent, of laminar structures which eventually 1048 01:06:47,010 --> 01:06:49,353 become turbulent and then they are modified, they are 1049 01:06:49,353 --> 01:06:51,680 transported. So we have the same problems and 1050 01:06:51,680 --> 01:06:54,760 we have to encode these causalities because trying to 1051 01:06:54,760 --> 01:06:58,560 learn all the physics, all the chemistry, all the biology from 1052 01:06:58,560 --> 01:07:02,640 scratch, using brute force training on data, it's not 1053 01:07:02,640 --> 01:07:04,760 efficient because we know the physics. 1054 01:07:04,960 --> 01:07:06,800 If we know the physics, let's use it. 1055 01:07:07,160 --> 01:07:10,800 But the problem is then how do you really encode that physics? 1056 01:07:10,800 --> 01:07:12,200 First of all, you need to know it. 1057 01:07:12,480 --> 01:07:16,280 So it's not the computer science guy who is going to know which 1058 01:07:16,280 --> 01:07:20,520 are the constraints which are the most suitable for high 1059 01:07:20,520 --> 01:07:23,920 energy jets, right? So we need the communities to 1060 01:07:23,920 --> 01:07:26,960 work together and also some problems which have been solved 1061 01:07:26,960 --> 01:07:29,240 about causalities. I don't know in high energy in 1062 01:07:29,240 --> 01:07:32,622 the energy physics community can be useful for fluid mechanics 1063 01:07:32,622 --> 01:07:36,800 and cover city. For instance, I have talked with 1064 01:07:36,960 --> 01:07:41,520 people from plasma physics. So in some cases they have 1065 01:07:41,520 --> 01:07:43,560 Navier–Stokes plus the magnetic fields. 1066 01:07:43,560 --> 01:07:46,880 They have problems of closures because they cannot resolve all 1067 01:07:46,880 --> 01:07:49,840 the scales they need to coarse- grained models. 1068 01:07:49,840 --> 01:07:54,760 So they have closure problems, and maybe something of what we 1069 01:07:54,760 --> 01:07:57,920 are doing with turbulence modeling can be reused in that 1070 01:07:57,920 --> 01:07:59,400 field. I'm collaborating with 1071 01:07:59,400 --> 01:08:02,280 volcanologists because in volcanoes you need to 1072 01:08:02,280 --> 01:08:05,960 characterize the rheology of magmas and you cannot measure it 1073 01:08:06,040 --> 01:08:10,320 and you have simplified models which are not accurate enough. 1074 01:08:10,640 --> 01:08:14,454 So you can use the techniques we use for turbulence modelling 1075 01:08:14,454 --> 01:08:16,600 to improve the rheology of magmas. 1076 01:08:16,608 --> 01:08:20,819 So that's where it becomes exciting to share knowledge 1077 01:08:20,828 --> 01:08:24,787 across the, the different scientific communities instead 1078 01:08:24,796 --> 01:08:29,868 of having every guy in its isolated department reinvent 1079 01:08:29,877 --> 01:08:34,080 everything from scratch because we are reinventing the wheel 1080 01:08:34,087 --> 01:08:37,688 otherwise, you know, and that's how we are trying to, you know, 1081 01:08:37,688 --> 01:08:40,743 I'm trying to, to, to, to animate this community. 1082 01:08:40,752 --> 01:08:45,000 And we have a seminar now called the AI for Science. 1083 01:08:45,008 --> 01:08:49,247 And hopefully we are trying to apply to funding opportunities, 1084 01:08:49,247 --> 01:08:53,836 even if they are not funding, just the fact that we are around 1085 01:08:53,836 --> 01:08:56,712 the table and we are thinking about what we could do together. 1086 01:08:56,712 --> 01:08:57,880 It's already something which 1087 01:08:57,880 --> 01:09:01,520 helps building the community. And then what people start to 1088 01:09:01,520 --> 01:09:04,840 collaborate and I think it's beneficial to everybody. 1089 01:09:04,840 --> 01:09:06,600 That's, that's my hope at least. 1090 01:09:06,608 --> 01:09:07,754 No, it sounds. Fantastic. 1091 01:09:07,761 --> 01:09:12,542 And I totally agree there's lots of I like the ideas that these 1092 01:09:12,550 --> 01:09:15,120 multidisciplinary centres, particularly in the age of AI. 1093 01:09:15,120 --> 01:09:16,800 I think it makes complete sense. 1094 01:09:16,800 --> 01:09:20,920 And as you say as well helping be a bridge to I guess the 1095 01:09:20,920 --> 01:09:23,640 industry and startups and and and things like that. 1096 01:09:23,640 --> 01:09:25,365 That makes makes complete sense. 1097 01:09:25,372 --> 01:09:30,562 I guess maybe to, as we come to the end, all of these new 1098 01:09:30,569 --> 01:09:34,746 things, all of these topics that have come up since, since I 1099 01:09:34,754 --> 01:09:39,344 guess you did your original studies, you know, how does this 1100 01:09:39,344 --> 01:09:44,303 affect the education system? You know, what should young 1101 01:09:44,310 --> 01:09:49,466 fluid mechanics researchers be be picking today as their PhD 1102 01:09:49,474 --> 01:09:53,768 topics or the undergraduate what what will prepare them for the 1103 01:09:53,776 --> 01:09:58,760 next 10 to 20 years of of the world we live in? 1104 01:09:58,768 --> 01:10:02,084 Yeah, I. Think we are in a very critical 1105 01:10:02,084 --> 01:10:05,842 period for education because well you and me have been 1106 01:10:05,850 --> 01:10:12,328 educated in a few without AI. So we have been taught how to to 1107 01:10:12,328 --> 01:10:19,336 solve problems without the help of AI to how to like to 1108 01:10:19,344 --> 01:10:24,796 interpret critically the results and how to look for information 1109 01:10:24,796 --> 01:10:29,360 and, you know, how to, to check if what we are doing is good or 1110 01:10:29,360 --> 01:10:32,336 not and so on. The problem is that the new 1111 01:10:32,344 --> 01:10:38,240 generation is AI native. So I, I was hearing, yeah, yeah. 1112 01:10:38,240 --> 01:10:41,440 Because yesterday at the radio 1113 01:10:41,440 --> 01:10:46,320 were saying that many teenagers now who are going to vote for 1114 01:10:46,320 --> 01:10:50,400 the first time are asking AIs what should they vote? 1115 01:10:50,800 --> 01:10:55,520 Because trust AI is more than politicians or more than the 1116 01:10:55,520 --> 01:10:57,200 classical media. OK. 1117 01:10:57,600 --> 01:11:00,440 And they're doing a little bit the same with the with the 1118 01:11:00,440 --> 01:11:04,520 school, with school teachers and with university professor who 1119 01:11:04,520 --> 01:11:07,920 are not prepared to that because, you know, the students 1120 01:11:07,920 --> 01:11:11,680 have smarter than we are and they know how to, you know, to 1121 01:11:11,680 --> 01:11:14,920 use AI to solve their fluid mechanics problems, to solve 1122 01:11:14,920 --> 01:11:19,160 their thermodynamics problems. And you have still teachers who 1123 01:11:20,160 --> 01:11:24,080 six months ago, we're not even imagining that the problem had 1124 01:11:24,080 --> 01:11:27,360 been solved by an AI. So right now, many, many 1125 01:11:27,360 --> 01:11:31,160 colleagues are scared about the fact that students are cheating 1126 01:11:31,200 --> 01:11:33,680 with AI. But the for me, the problem is 1127 01:11:33,680 --> 01:11:36,640 not only cheating. The fact is that some students 1128 01:11:36,640 --> 01:11:42,040 are basically, you know, delegating the task of thinking 1129 01:11:42,040 --> 01:11:44,360 to AI. That's more much more dangerous. 1130 01:11:44,720 --> 01:11:49,760 So what I think is that we have to rethink the whole system and 1131 01:11:49,760 --> 01:11:54,000 many tasks which were, you know, I teach you how to solve a 1132 01:11:54,000 --> 01:11:59,000 second order linear ordinary differential equations and you 1133 01:11:59,000 --> 01:12:01,000 just have the technique to solve that. 1134 01:12:01,000 --> 01:12:05,080 That's no longer useful because you know, ChatGPT knows to do 1135 01:12:05,080 --> 01:12:08,280 that very well. The problem, we should teach the 1136 01:12:08,280 --> 01:12:12,800 students how to analyze the results and how to understand if 1137 01:12:12,800 --> 01:12:18,880 the result is correct, if it's an hallucination, how to 1138 01:12:18,880 --> 01:12:22,000 formulate the questions also. Because if you don't know how to 1139 01:12:22,000 --> 01:12:27,720 ask AI correctly, what to do, AI is going to do something you 1140 01:12:27,720 --> 01:12:31,560 know you don't control. And if you are not cultivated 1141 01:12:31,560 --> 01:12:35,920 enough to understand the answer, you are just going to accept 1142 01:12:35,920 --> 01:12:41,600 whatever comes out without, without, you know, without 1143 01:12:41,680 --> 01:12:43,760 criticizing what what's coming out. 1144 01:12:44,160 --> 01:12:46,560 So I think that we have to move the shift. 1145 01:12:46,560 --> 01:12:50,480 So instead of teaching techniques and, you know, 1146 01:12:50,480 --> 01:12:57,080 techniques or, or just notions, we should teach critical 1147 01:12:57,080 --> 01:12:59,560 thinking. And that's harder because it, 1148 01:13:00,120 --> 01:13:05,320 it, it, it implies a, a considerable modification of our 1149 01:13:05,320 --> 01:13:07,480 teaching programs and new teaching approaches. 1150 01:13:07,760 --> 01:13:10,920 It's not easy to do, especially in large universities as we 1151 01:13:10,920 --> 01:13:16,560 have, I, we have courses with, you know, 500 students. 1152 01:13:16,840 --> 01:13:21,760 It's not easy to, you know, to have to, to change all the 1153 01:13:21,760 --> 01:13:24,560 teaching procedure. But it's something we need to do 1154 01:13:25,120 --> 01:13:29,680 because without critical thinking, basically humans are 1155 01:13:29,680 --> 01:13:35,920 going to lose competence. And yeah, and, and, and, and, 1156 01:13:36,280 --> 01:13:39,920 and in case, in case AI is not working, what do we do? 1157 01:13:40,240 --> 01:13:44,520 And I had a very nice metaphor from a person from the European 1158 01:13:44,520 --> 01:13:50,800 committee who said, well, when you, when you are, when are you 1159 01:13:50,800 --> 01:13:58,040 are a pilot of an aircraft, you are 9, I say 8000 or 99900. 1160 01:13:58,040 --> 01:14:01,548 Ninety, 100% of the time you are going to use the automatic 1161 01:14:01,548 --> 01:14:04,455 pilot. But in case of accident, the 1162 01:14:04,463 --> 01:14:10,512 pilot needs to know how to drive the plane down to safety, right? 1163 01:14:10,512 --> 01:14:12,640 And that's exactly that. 1164 01:14:12,640 --> 01:14:17,720 So we, we need to be critical because we don't know where the 1165 01:14:17,720 --> 01:14:20,680 AI is going to hallucinate even we are. 1166 01:14:20,720 --> 01:14:25,040 Also, we need to, to understand if, for instance, sometimes AI 1167 01:14:25,040 --> 01:14:28,634 is able to bring us notions from other disciplines which are 1168 01:14:28,634 --> 01:14:32,160 not our discipline. But if we have methodology, we 1169 01:14:32,160 --> 01:14:38,080 know how to analyze if what the, the, the, the LLM is saying is 1170 01:14:38,080 --> 01:14:41,680 reasonable or not. We know how to verify facts and 1171 01:14:41,680 --> 01:14:45,680 so on. And we have culture, general 1172 01:14:45,680 --> 01:14:50,680 culture and language capacity to understand then we can control 1173 01:14:50,680 --> 01:14:54,680 and we can be enriched as professionals, as scientists, as 1174 01:14:54,680 --> 01:14:59,160 humans and whatever instead of being just, you know, replaced 1175 01:14:59,520 --> 01:15:01,920 like in Matrix. That's my, my nightmare. 1176 01:15:01,920 --> 01:15:05,480 You know, Matrix humans are just there to be pumped energy. 1177 01:15:08,960 --> 01:15:09,480 Yeah. No. 1178 01:15:09,480 --> 01:15:14,040 It does. It does seem that it's not 1179 01:15:14,040 --> 01:15:16,600 universally accepted. How best to do this from an 1180 01:15:16,600 --> 01:15:19,960 education point of view? I've heard some go the other way 1181 01:15:19,960 --> 01:15:24,000 and be almost well for an undergraduate, we will do lots 1182 01:15:24,000 --> 01:15:29,720 of in person examinations and basically no homework anymore 1183 01:15:29,760 --> 01:15:34,080 because you this is no way of knowing the homework. 1184 01:15:35,480 --> 01:15:39,680 And, and as you say, there's the other side, which is maybe a 1185 01:15:39,680 --> 01:15:45,760 little bit more ex, you know, do a presentation to explain how 1186 01:15:45,760 --> 01:15:49,000 you got there. I, I kind of lean a little bit 1187 01:15:49,000 --> 01:15:52,400 more to the first one at the right stage of your career. 1188 01:15:52,840 --> 01:15:55,905 I guess this is more like even a secondary school, let's say, 1189 01:15:55,905 --> 01:15:59,580 you know, before university, you know, the certain things where 1190 01:15:59,580 --> 01:16:03,880 I just feel you have to know it because as you say, what if 1191 01:16:03,880 --> 01:16:06,560 there is a moment where you don't, you're on a, you know, 1192 01:16:06,560 --> 01:16:08,280 you don't have access to the AI tools. 1193 01:16:09,480 --> 01:16:13,200 I feel there's like a transition point when you can expect. 1194 01:16:14,040 --> 01:16:15,680 Like if you're doing a P. HD. 1195 01:16:17,000 --> 01:16:18,400 Or certainly a. Postdoc. 1196 01:16:18,480 --> 01:16:22,400 I actually think AI is a fantastic help and will make you 1197 01:16:22,400 --> 01:16:26,640 more productive, make you be able to do more things, help you 1198 01:16:26,640 --> 01:16:28,960 achieve what you want to do faster and better, and 1199 01:16:28,960 --> 01:16:32,560 ultimately get a, you know what would have taken you a month to 1200 01:16:32,560 --> 01:16:36,440 write some coding, which wasn't your main task, it was just you 1201 01:16:36,440 --> 01:16:39,480 had to do it to get there. AI will accelerate it. 1202 01:16:39,480 --> 01:16:43,240 But if you're at the undergraduate level or or still 1203 01:16:43,240 --> 01:16:48,240 at school or something, I feel that's probably where AI should 1204 01:16:48,240 --> 01:16:53,320 only be used to help you learn something, but not it shouldn't 1205 01:16:53,320 --> 01:16:57,120 be used like a calculator almost or or or you will. 1206 01:16:57,120 --> 01:16:58,760 Just not. Understand the fundamental 1207 01:16:59,320 --> 01:17:04,120 theory, it can be used. As a mate you know because you 1208 01:17:04,120 --> 01:17:11,560 can no say tell the AI ask me questions and give the answer 1209 01:17:11,560 --> 01:17:14,000 and then the AI can analyze the tutor. 1210 01:17:14,240 --> 01:17:18,000 Yes, and it. Makes, you know, learning more 1211 01:17:18,000 --> 01:17:20,200 interactive. In my research, sometimes I feel 1212 01:17:20,200 --> 01:17:23,200 alone and I don't have a colleague to discuss about the 1213 01:17:23,200 --> 01:17:25,440 point. So I take an AI and say, what do 1214 01:17:25,440 --> 01:17:28,480 you think about this idea? And so we start the conversation 1215 01:17:28,480 --> 01:17:31,080 and that's useful. And then, you know, the AI is 1216 01:17:31,080 --> 01:17:33,120 bringing some ideas. So, yeah, yeah. 1217 01:17:33,120 --> 01:17:35,000 Or. Explaining new topic. 1218 01:17:35,000 --> 01:17:38,040 I often do that when I'm trying to and you can ask it. 1219 01:17:38,040 --> 01:17:40,440 And the great thing I find is that it has these different 1220 01:17:40,440 --> 01:17:44,144 levels where you could say, OK, sometimes I'll ask like, can 1221 01:17:44,144 --> 01:17:47,498 you show me that in a code like how would you actually code 1222 01:17:47,498 --> 01:17:50,581 this up? Which would be hard for a tutor 1223 01:17:50,581 --> 01:17:56,520 to do instantly like a person to be that flexible to move from 1224 01:17:56,520 --> 01:17:59,290 reasoning. To code and back and so on. 1225 01:17:59,298 --> 01:18:02,732 Because you are, yeah, easily lost in, you know, coding 1226 01:18:02,740 --> 01:18:06,776 details and so on That that yeah, that's that's actually 1227 01:18:06,784 --> 01:18:10,388 great. But the problem is to explain to 1228 01:18:10,388 --> 01:18:14,242 the young generations that that's the way probably to use 1229 01:18:14,250 --> 01:18:19,360 the AI more like a companion then like an entity that is 1230 01:18:19,368 --> 01:18:26,224 going to do the work in my stead while I'm, I don't know, playing 1231 01:18:26,224 --> 01:18:30,320 video games or something. That's that's not the point. 1232 01:18:30,328 --> 01:18:34,844 And, and they need so the problem is to explain them what 1233 01:18:34,844 --> 01:18:40,308 are the risks and why they are missing something if they use AI 1234 01:18:40,308 --> 01:18:43,156 only for that. So I can understand that 1235 01:18:43,164 --> 01:18:45,190 sometimes you need to accelerate, you know, you have 1236 01:18:45,198 --> 01:18:48,200 to give a report or you are a little bit in a hurry. 1237 01:18:48,208 --> 01:18:50,670 But if you do that systematically, then basically 1238 01:18:50,678 --> 01:18:54,255 you are missing your education because it's not you who is 1239 01:18:54,263 --> 01:18:57,900 going to be educated, but the AI and and I and I think. 1240 01:18:57,908 --> 01:19:02,362 In some ways it'll end up creating a bit of a two-tier 1241 01:19:02,370 --> 01:19:07,506 system where in some ways people will become it's quite capable 1242 01:19:07,514 --> 01:19:13,282 of doing some jobs where it requires them just to be, you 1243 01:19:13,290 --> 01:19:16,565 know, a functional user of these tools. 1244 01:19:16,573 --> 01:19:22,680 But it will stop you from maybe being the the inventor or the 1245 01:19:22,688 --> 01:19:24,878 innovator. And there it there feels like 1246 01:19:24,886 --> 01:19:28,674 you could, to be fair, you could probably still do quite good at 1247 01:19:28,674 --> 01:19:32,028 a job if you because these AI tools are so good. 1248 01:19:32,036 --> 01:19:36,400 But I feel there must be some points where that lack of deep 1249 01:19:36,408 --> 01:19:39,075 understanding will come and bite you. 1250 01:19:39,083 --> 01:19:44,780 And so, yeah, it's it's a challenge. 1251 01:19:44,788 --> 01:19:47,320 Yeah. We're not in very. 1252 01:19:47,328 --> 01:19:51,274 Specialistic fields as as as so we are in very specialistic 1253 01:19:51,282 --> 01:19:55,113 fields. So I don't think that the AI has 1254 01:19:55,113 --> 01:19:58,842 been trained enough. So that's where you know, you 1255 01:19:58,850 --> 01:20:03,392 bring the new ideas, maybe the AI strengthens you, accelerates 1256 01:20:03,392 --> 01:20:08,560 you with the coding or allows you to make connections more 1257 01:20:08,568 --> 01:20:11,540 quickly and so on. But it's still new. 1258 01:20:11,548 --> 01:20:16,886 We're bringing the ideas. And if you're you know although 1259 01:20:16,886 --> 01:20:22,229 that. I do suspect will also be, you 1260 01:20:22,237 --> 01:20:28,415 know, succeeded by AI because the, you know, the, the AI 1261 01:20:28,423 --> 01:20:33,956 scientist or the AI engineer. I have to say, I, I, I still 1262 01:20:33,964 --> 01:20:37,704 suspect actually that it will be able to come up with as many 1263 01:20:37,712 --> 01:20:40,766 ideas as as we, we will be able to do. 1264 01:20:40,774 --> 01:20:45,396 Maybe not the, you know, Einstein new completely 1265 01:20:45,404 --> 01:20:50,730 breakthrough thing, but it's it's certainly going to be an 1266 01:20:50,738 --> 01:20:53,010 interesting test. Of given. 1267 01:20:53,018 --> 01:20:56,556 That the whole, like your centre is a good example of it. 1268 01:20:56,564 --> 01:21:00,420 I, I do get the sense that AI for science or AI for 1269 01:21:00,428 --> 01:21:02,474 engineering is now becoming the frontier. 1270 01:21:02,482 --> 01:21:06,916 It wasn't five years ago. It was sort of seen as a niche 1271 01:21:06,924 --> 01:21:10,189 area. I, I get the sense now that the, 1272 01:21:10,189 --> 01:21:13,962 the world of sort of physical AI and AI for engineering and AI 1273 01:21:13,962 --> 01:21:16,046 for science is becoming a hot topic. 1274 01:21:16,054 --> 01:21:19,520 And whenever it's a hot topic and there's lots of investment 1275 01:21:19,528 --> 01:21:22,240 and money put into it, it, you know, yeah. 1276 01:21:22,248 --> 01:21:25,486 There is, yeah, some overselling, yeah. 1277 01:21:25,494 --> 01:21:30,652 But but yes, I guess, I guess that's the frontier. 1278 01:21:30,660 --> 01:21:33,486 And I, well, I want to believe that. 1279 01:21:33,494 --> 01:21:37,240 Well, even if the guys are becoming more and more powerful 1280 01:21:37,240 --> 01:21:41,905 and potentially they could even win a Nobel Prize at some point 1281 01:21:41,905 --> 01:21:45,300 if you think about AlphaFold for instance. 1282 01:21:45,308 --> 01:21:50,265 It's well. What's new is the fact that in a 1283 01:21:50,265 --> 01:21:53,335 few years they could discover thousands and hundreds of 1284 01:21:53,343 --> 01:21:57,608 thousands of proteins which are much more than the 2000 and 1285 01:21:57,616 --> 01:22:01,114 something proteins they had discovered during the past 50 1286 01:22:01,122 --> 01:22:05,300 years. So every protein structure was 1287 01:22:05,308 --> 01:22:10,595 PhD thesis and now you can generate plenty of them, you 1288 01:22:10,603 --> 01:22:13,196 know, just with with with these AI. 1289 01:22:13,204 --> 01:22:19,034 So what was new was the fact of being able to accelerate so 1290 01:22:19,042 --> 01:22:21,260 much. But still without the knowledge 1291 01:22:21,260 --> 01:22:25,005 of the proteins from the past, it wouldn't have been possible 1292 01:22:25,013 --> 01:22:27,638 to have that. And the same for climate. 1293 01:22:27,646 --> 01:22:32,768 If now climate models are so good, that's because for 10s of 1294 01:22:32,768 --> 01:22:36,920 years people have developed better and better climate model, 1295 01:22:36,920 --> 01:22:41,224 the classical ones, and they have done data simulation and 1296 01:22:41,232 --> 01:22:45,192 then they have done all the reanalysis from the beginning. 1297 01:22:45,200 --> 01:22:48,191 And with that they have generated the huge databases 1298 01:22:48,199 --> 01:22:50,900 which can be used for by the models. 1299 01:22:50,908 --> 01:22:53,664 And also, well, everything is like that. 1300 01:22:53,672 --> 01:22:57,880 That's because we have huge amounts of human knowledge 1301 01:22:57,888 --> 01:23:02,802 accumulated over years, which can be injected in these models. 1302 01:23:02,802 --> 01:23:05,480 But what once all the knowledge 1303 01:23:05,480 --> 01:23:07,880 will be generated by by these things? 1304 01:23:08,160 --> 01:23:11,080 Probably, I don't know, Maybe at some point they will become so 1305 01:23:11,080 --> 01:23:14,840 intelligent with these emergent phenomena, they will be able to 1306 01:23:14,840 --> 01:23:18,280 create something new that happens in very complex systems. 1307 01:23:19,200 --> 01:23:24,360 But I, I, I I I'm. But I, I hope humans has still a 1308 01:23:24,360 --> 01:23:26,160 role to play. Oh yes. 1309 01:23:27,040 --> 01:23:29,840 And that. Well, and that instead of being 1310 01:23:29,840 --> 01:23:34,531 replaced by these AIs, well, one of the person who 1311 01:23:34,531 --> 01:23:37,610 participates, a Jesse Thaler, he's a, he's a physicist from 1312 01:23:37,610 --> 01:23:41,640 MIT and he participated to our AI for science launch day. 1313 01:23:41,920 --> 01:23:45,560 And he came out with his metaphor of the Centaur 1314 01:23:45,680 --> 01:23:48,720 scientists. And I found that as a beautiful 1315 01:23:48,720 --> 01:23:53,960 metaphor because as you know, if you are two men and not enough 1316 01:23:53,960 --> 01:23:57,360 horse, you are too slow. If you are two horse and not 1317 01:23:57,360 --> 01:24:02,160 enough man, you cannot reason. But if you have 1/2 man half 1318 01:24:02,160 --> 01:24:05,360 horse, you can go as fast as a horse while having all, you 1319 01:24:05,360 --> 01:24:08,840 know, the creativity and and width of a human. 1320 01:24:09,160 --> 01:24:13,800 And so, well, I, I adopted this idea and say that, well, our 1321 01:24:13,880 --> 01:24:17,760 role would be to generate to, to, to train the next generation 1322 01:24:17,760 --> 01:24:22,000 of Centaur scientists and Centaur professionals and not 1323 01:24:22,000 --> 01:24:26,320 just horse. I like that. 1324 01:24:26,640 --> 01:24:29,480 You know, and that's, that's probably a great way to end this 1325 01:24:29,480 --> 01:24:32,600 in the, in the sense of I agree with you that it's, it's about 1326 01:24:33,040 --> 01:24:36,320 the coupling of, of human and AI. 1327 01:24:36,320 --> 01:24:41,920 It's about being progressive and optimistic about AI, but not 1328 01:24:41,920 --> 01:24:45,800 forgetting the, the sort of human role role in it. 1329 01:24:45,800 --> 01:24:50,480 And, and I would argue that today the human specialist is 1330 01:24:50,480 --> 01:24:54,760 needed even more because, you know, to, to help develop these 1331 01:24:54,760 --> 01:24:57,080 AI models and make them accurate and point them in the right 1332 01:24:57,080 --> 01:25:00,080 direction. But maybe we need to talk again 1333 01:25:00,080 --> 01:25:04,320 in two years time and, and see where things have gone. 1334 01:25:04,320 --> 01:25:07,775 It'd be interesting to to listen back and see as it shot off 1335 01:25:07,775 --> 01:25:10,160 like this is, is it the same? Is it less? 1336 01:25:10,200 --> 01:25:13,219 That will be an interesting, but I really enjoyed this 1337 01:25:13,219 --> 01:25:15,816 discussion with you and I personally always love working 1338 01:25:15,816 --> 01:25:18,448 with you on on various committees and things. 1339 01:25:18,456 --> 01:25:21,949 And yeah, been a pleasure to have to spend this time with 1340 01:25:21,957 --> 01:25:23,454 you. OK. 1341 01:25:23,462 --> 01:25:26,840 Thank you for the interview. It was very, very nice to 1342 01:25:26,848 --> 01:25:29,134 discuss about all this. Great. 1343 01:25:29,142 --> 01:25:29,640 Thank you.