Large physics models boost the performance of passenger cars

WEBVTT
1
00:00:08.000 --> 00:00:10.960
- Hello, and welcome to the Physics World weekly
2
00:00:10.960 --> 00:00:13.059
- podcast. I'm Hamish Johnston.
3
00:00:13.679 --> 00:00:15.779
- My guest is David Wheater,
4
00:00:16.295 --> 00:00:17.675
- who is vice president
5
00:00:18.054 --> 00:00:18.554
- automotive
6
00:00:19.175 --> 00:00:22.155
- at UK and US based PhysicsX.
7
00:00:23.574 --> 00:00:24.394
- An aeronautical
8
00:00:24.855 --> 00:00:27.734
- engineer by training, Dave spent much of his
9
00:00:27.734 --> 00:00:29.515
- career using aerodynamics
10
00:00:29.975 --> 00:00:31.355
- to improve the performance
11
00:00:31.940 --> 00:00:35.719
- of formula one racing cars for several different
12
00:00:35.859 --> 00:00:37.079
- f one constructors.
13
00:00:38.020 --> 00:00:40.520
- Now his focus is on using
14
00:00:40.899 --> 00:00:43.399
- AI based large physics models
15
00:00:43.780 --> 00:00:47.255
- to design better passenger cars and other vehicles.
16
00:00:48.034 --> 00:00:49.414
- This involves combining
17
00:00:49.954 --> 00:00:50.454
- computational
18
00:00:50.835 --> 00:00:51.655
- fluid dynamics
19
00:00:51.954 --> 00:00:52.454
- simulations
20
00:00:53.234 --> 00:00:57.254
- with wind tunnel experiments and real world testing.
21
00:00:57.715 --> 00:01:01.254
- As Dave explains in this wide ranging conversation,
22
00:01:01.634 --> 00:01:04.269
- which also looks at career opportunities
23
00:01:04.729 --> 00:01:05.629
- for physicists
24
00:01:06.250 --> 00:01:07.469
- in this multidisciplinary
25
00:01:08.329 --> 00:01:08.829
- field.
26
00:01:16.115 --> 00:01:18.055
- Hi, Dave. Welcome to the podcast.
27
00:01:19.314 --> 00:01:21.155
- Hi, Hamish. Thank you very much for having
28
00:01:21.155 --> 00:01:21.655
- me.
29
00:01:22.195 --> 00:01:24.435
- So Dave, can we start out? Can you
30
00:01:24.435 --> 00:01:26.295
- tell us a bit about yourself?
31
00:01:27.954 --> 00:01:30.034
- Who who are you? What's your background? And,
32
00:01:31.520 --> 00:01:33.939
- what's your experience in the automotive industry?
33
00:01:35.520 --> 00:01:36.020
- Sure.
34
00:01:36.560 --> 00:01:40.159
- So I am technical director for automotive at
35
00:01:40.159 --> 00:01:40.659
- PhysX.
36
00:01:41.600 --> 00:01:44.240
- That means I I lead automotive activities and
37
00:01:44.240 --> 00:01:45.859
- also competition engineering,
38
00:01:47.055 --> 00:01:48.114
- based on my background.
39
00:01:49.055 --> 00:01:50.674
- So what does that involve?
40
00:01:51.055 --> 00:01:52.114
- Working with customers,
41
00:01:52.655 --> 00:01:53.474
- on the ground.
42
00:01:54.734 --> 00:01:56.674
- You know, we we co engineer
43
00:01:57.055 --> 00:01:59.375
- our our solutions with with customers in order
44
00:01:59.375 --> 00:02:00.515
- to kind of quickly,
45
00:02:01.329 --> 00:02:02.709
- build value. We have
46
00:02:03.090 --> 00:02:06.469
- experienced engineering teams, cross functional teams across
47
00:02:06.770 --> 00:02:10.050
- engineering, machine learning, data science, and software. And,
48
00:02:10.050 --> 00:02:10.870
- you know, living
49
00:02:11.250 --> 00:02:13.349
- side by side with our customers on-site,
50
00:02:13.854 --> 00:02:15.534
- you know, we we we really learn about
51
00:02:15.534 --> 00:02:17.634
- the real world of our customers and,
52
00:02:18.415 --> 00:02:19.615
- and and what we need to do in
53
00:02:19.615 --> 00:02:20.514
- the hands of engineers.
54
00:02:20.895 --> 00:02:22.675
- So I I work there, and I also
55
00:02:23.055 --> 00:02:25.455
- coordinate across the other key pillars of our
56
00:02:25.455 --> 00:02:25.955
- company
57
00:02:26.335 --> 00:02:26.995
- for automotive.
58
00:02:27.689 --> 00:02:29.849
- We have a research group. We'll talk later
59
00:02:29.849 --> 00:02:31.870
- about their work on the large physics model.
60
00:02:32.090 --> 00:02:34.430
- But, yeah, that's a really world class group,
61
00:02:34.810 --> 00:02:37.849
- contributing to academia. For example, a spotlight paper
62
00:02:37.849 --> 00:02:38.509
- at NeurIPS,
63
00:02:38.810 --> 00:02:40.750
- one of the leading machine learning conferences.
64
00:02:41.764 --> 00:02:43.525
- Also, you know, keeping us at the cutting
65
00:02:43.525 --> 00:02:45.444
- edge of research for the benefit of our
66
00:02:45.444 --> 00:02:45.944
- customers.
67
00:02:46.805 --> 00:02:48.025
- Recent example there,
68
00:02:48.724 --> 00:02:51.525
- well, very recent, in fact, Transolver three, which
69
00:02:51.525 --> 00:02:53.865
- is a a kind of machine learning architecture
70
00:02:54.245 --> 00:02:55.319
- similar to a,
71
00:02:56.040 --> 00:02:57.099
- sort of transformer
72
00:02:57.480 --> 00:03:00.540
- used in large languages but architected for physics.
73
00:03:01.159 --> 00:03:01.900
- That was,
74
00:03:02.840 --> 00:03:05.480
- published in academia late February, and we're already
75
00:03:05.480 --> 00:03:07.260
- shipping that onto our platform.
76
00:03:07.835 --> 00:03:10.235
- And then finally, you know, we've got an
77
00:03:10.235 --> 00:03:10.735
- amazing,
78
00:03:11.435 --> 00:03:12.974
- AI and and and product
79
00:03:13.275 --> 00:03:15.915
- group who who build a platform that that
80
00:03:15.915 --> 00:03:17.995
- everything is based on and that allows us
81
00:03:17.995 --> 00:03:20.895
- to deploy our end to end AI workflows
82
00:03:21.034 --> 00:03:22.014
- into our customers'
83
00:03:22.590 --> 00:03:24.689
- most most difficult physics problems.
84
00:03:25.790 --> 00:03:27.389
- And and and there, you know, we need
85
00:03:27.389 --> 00:03:29.310
- to make sure that we're integrating with the
86
00:03:29.310 --> 00:03:31.389
- way our customers already work. So a lot
87
00:03:31.389 --> 00:03:32.290
- of work around,
88
00:03:34.030 --> 00:03:37.069
- kind of hook hooks into existing simulation and
89
00:03:37.069 --> 00:03:37.730
- CAD tools
90
00:03:38.645 --> 00:03:40.824
- and the right level accessibility for engineers.
91
00:03:41.125 --> 00:03:43.064
- So, you know, my my work,
92
00:03:43.604 --> 00:03:46.504
- the customers and understanding the customer problems is
93
00:03:46.805 --> 00:03:48.884
- is really useful for for working with those
94
00:03:48.884 --> 00:03:49.384
- groups.
95
00:03:50.245 --> 00:03:52.264
- I should say I've I've been at PhysicsX
96
00:03:52.324 --> 00:03:53.305
- for nine months,
97
00:03:54.389 --> 00:03:56.069
- and and and it's been, you know, an
98
00:03:56.069 --> 00:03:56.969
- amazing experience,
99
00:03:57.829 --> 00:04:00.329
- working with that kind of diverse group.
100
00:04:01.109 --> 00:04:03.430
- And and prior to PhysicsX, I I I
101
00:04:03.430 --> 00:04:05.269
- spent quite a few years, twenty five years
102
00:04:05.269 --> 00:04:06.405
- in Formula One,
103
00:04:07.365 --> 00:04:09.064
- starting my life as an aerodynamicist
104
00:04:09.685 --> 00:04:10.584
- at at Benetton.
105
00:04:11.764 --> 00:04:12.264
- And,
106
00:04:12.724 --> 00:04:14.965
- by the mid two thousands, I was leading
107
00:04:14.965 --> 00:04:15.625
- the aerodynamics
108
00:04:15.925 --> 00:04:16.425
- development
109
00:04:16.805 --> 00:04:19.064
- across half the car for Renault f one.
110
00:04:19.650 --> 00:04:20.850
- You know, that was still in the days
111
00:04:20.850 --> 00:04:23.430
- of fairly small aerodynamics departments, but
112
00:04:23.810 --> 00:04:24.790
- we were an innovative
113
00:04:25.170 --> 00:04:26.310
- group and and launched
114
00:04:26.850 --> 00:04:27.730
- 2005
115
00:04:27.730 --> 00:04:29.509
- with with some pretty novel ideas.
116
00:04:30.050 --> 00:04:33.009
- One of which, coincidentally, was developed by our
117
00:04:33.009 --> 00:04:33.509
- cofounder,
118
00:04:33.875 --> 00:04:36.134
- Robin Talui, which was a a mass damper.
119
00:04:37.154 --> 00:04:39.014
- And I worked on a
120
00:04:39.395 --> 00:04:42.035
- a a novel rear suspension system, which we
121
00:04:42.035 --> 00:04:44.514
- introduced that year to replicate the error effect
122
00:04:44.514 --> 00:04:45.955
- of of some of the floor that had
123
00:04:45.955 --> 00:04:48.055
- been removed due to regulation changes.
124
00:04:49.300 --> 00:04:51.400
- And and that really kind of cemented,
125
00:04:51.779 --> 00:04:52.520
- the championship
126
00:04:52.900 --> 00:04:54.180
- win in 2005
127
00:04:54.180 --> 00:04:55.220
- and 2006,
128
00:04:55.220 --> 00:04:56.980
- and that was at a time when you
129
00:04:56.980 --> 00:04:59.560
- could win championships with with novel ideas.
130
00:05:00.020 --> 00:05:01.939
- If I jump forward to my later life
131
00:05:01.939 --> 00:05:03.014
- in Formula one,
132
00:05:03.574 --> 00:05:06.694
- 2018, I became aerodynamics director at Williams and
133
00:05:06.694 --> 00:05:07.194
- then,
134
00:05:07.814 --> 00:05:09.915
- aerodynamics technical director at Alpine.
135
00:05:10.455 --> 00:05:13.035
- Much larger departments by then, real machines
136
00:05:13.415 --> 00:05:14.395
- of of development
137
00:05:14.855 --> 00:05:18.155
- where winning takes just relentless pursuit of performance
138
00:05:18.375 --> 00:05:18.990
- over time.
139
00:05:20.430 --> 00:05:22.589
- And and that's where I I met PhysicsX,
140
00:05:22.589 --> 00:05:24.509
- and and they stepped into the picture with
141
00:05:24.509 --> 00:05:25.569
- me as a customer.
142
00:05:27.310 --> 00:05:29.709
- We did some exciting work together. I can't
143
00:05:29.709 --> 00:05:31.709
- talk about the details, but, yeah, the the
144
00:05:31.709 --> 00:05:32.209
- impact
145
00:05:32.589 --> 00:05:34.050
- was what caught my attention,
146
00:05:34.995 --> 00:05:36.834
- not just that we we could do things
147
00:05:36.834 --> 00:05:37.495
- much faster,
148
00:05:38.274 --> 00:05:40.595
- but but the real transformation came from what
149
00:05:40.595 --> 00:05:42.774
- we could do differently with that speed.
150
00:05:43.555 --> 00:05:45.254
- In our case, reacting to results,
151
00:05:45.954 --> 00:05:47.314
- over a race weekend in a in a
152
00:05:47.314 --> 00:05:49.334
- way that just wasn't possible before.
153
00:05:49.759 --> 00:05:51.519
- And and, David, I I think it's easy
154
00:05:51.519 --> 00:05:54.339
- for a physicist to imagine how aerodynamics
155
00:05:54.800 --> 00:05:58.399
- can be very important in Formula one. But
156
00:05:58.399 --> 00:06:01.199
- what about the design of of of modern
157
00:06:01.199 --> 00:06:02.959
- cars, you know, the sort of car that
158
00:06:02.959 --> 00:06:04.979
- I would drive to the supermarket?
159
00:06:05.865 --> 00:06:09.944
- Can significant improvements in fuel efficiency, handling, and
160
00:06:09.944 --> 00:06:10.444
- safety
161
00:06:10.824 --> 00:06:13.805
- still be gained by improving aerodynamics?
162
00:06:15.704 --> 00:06:18.024
- Yeah. Absolutely. And and and as you say,
163
00:06:18.024 --> 00:06:20.550
- Hamish, you know, in in in Formula one,
164
00:06:20.709 --> 00:06:23.449
- aerodynamics and downforce is is is the currency.
165
00:06:23.589 --> 00:06:24.789
- You know, 11%
166
00:06:24.789 --> 00:06:26.870
- of downforce is a tenth of a second,
167
00:06:26.870 --> 00:06:28.569
- and and that can be a a championship
168
00:06:28.629 --> 00:06:30.550
- changer for for a team in a tight
169
00:06:30.550 --> 00:06:31.050
- championship.
170
00:06:32.550 --> 00:06:34.069
- And and I should say in, you know,
171
00:06:34.069 --> 00:06:34.514
- in,
172
00:06:35.074 --> 00:06:37.095
- Formula One, it's not just about the downforce.
173
00:06:37.795 --> 00:06:40.035
- It's also about the drivability. You know, drivers
174
00:06:40.115 --> 00:06:42.275
- but they don't drive to the peak. They
175
00:06:42.275 --> 00:06:44.295
- drive to the trough if you've got variations
176
00:06:44.435 --> 00:06:46.295
- in downforce over over car conditions.
177
00:06:47.699 --> 00:06:48.439
- And so
178
00:06:49.300 --> 00:06:50.120
- in automotive,
179
00:06:51.539 --> 00:06:52.839
- kind of passenger vehicles,
180
00:06:53.300 --> 00:06:56.099
- aerodynamics plays just as critical a role in
181
00:06:56.099 --> 00:06:56.919
- in in slightly
182
00:06:57.459 --> 00:06:58.360
- different ways,
183
00:06:59.300 --> 00:07:01.495
- and gets into just as many systems on
184
00:07:01.495 --> 00:07:03.334
- the car. The obvious one being, of course,
185
00:07:03.334 --> 00:07:06.454
- the external shape, the focus there more on
186
00:07:06.454 --> 00:07:08.954
- drag than than downforce from Formula one.
187
00:07:09.334 --> 00:07:11.514
- You know, above 30 miles an hour,
188
00:07:12.149 --> 00:07:14.230
- most of most of your energy consumption is
189
00:07:14.230 --> 00:07:16.649
- is going into overcoming air resistance. So
190
00:07:16.949 --> 00:07:17.930
- good fuel efficiency
191
00:07:18.230 --> 00:07:21.430
- absolutely needs great aerodynamics, whether that's kind of
192
00:07:21.430 --> 00:07:23.269
- the obvious details that you see on the
193
00:07:23.269 --> 00:07:24.649
- car, the spoilers, etcetera,
194
00:07:24.995 --> 00:07:26.675
- or or the lot less obvious like the
195
00:07:26.675 --> 00:07:28.754
- tire shapes, which can can be real really
196
00:07:28.754 --> 00:07:29.254
- important.
197
00:07:30.115 --> 00:07:31.814
- But also in the thermal management,
198
00:07:32.274 --> 00:07:33.415
- systems like powertrains,
199
00:07:34.194 --> 00:07:37.314
- NVH, which is the noise vibration harshness, and
200
00:07:37.314 --> 00:07:39.875
- that's about controlling the turbulence of the flow
201
00:07:39.875 --> 00:07:40.680
- around the car,
202
00:07:41.240 --> 00:07:44.039
- and, of course, passenger comfort, HVAC, which is
203
00:07:44.039 --> 00:07:45.579
- heating, ventilation, and aircon.
204
00:07:46.199 --> 00:07:48.759
- Those are really important attributes for for, you
205
00:07:48.759 --> 00:07:50.860
- know, OEMs developing their cars,
206
00:07:51.479 --> 00:07:53.180
- and and, you know, pitching them,
207
00:07:54.120 --> 00:07:55.180
- at the top of the market.
208
00:07:56.865 --> 00:08:00.084
- And so today, how how are vehicle
209
00:08:00.384 --> 00:08:00.884
- aerodynamics
210
00:08:01.985 --> 00:08:02.485
- simulated?
211
00:08:03.024 --> 00:08:05.425
- Is it, is it sort of a I'm
212
00:08:05.425 --> 00:08:07.685
- guessing you're you're using some pretty powerful,
213
00:08:08.865 --> 00:08:09.845
- computer models,
214
00:08:10.699 --> 00:08:13.020
- to simulate how, you know, for example, how
215
00:08:13.020 --> 00:08:15.660
- air flows around a car, how it interacts
216
00:08:15.660 --> 00:08:17.759
- with the tires, that sort of thing?
217
00:08:19.500 --> 00:08:20.000
- Absolutely.
218
00:08:20.300 --> 00:08:20.699
- So,
219
00:08:22.060 --> 00:08:22.720
- you know,
220
00:08:24.785 --> 00:08:25.285
- traditionally,
221
00:08:26.545 --> 00:08:28.085
- over the last few decades,
222
00:08:28.785 --> 00:08:31.285
- we've been using these numerical simulations
223
00:08:31.905 --> 00:08:34.725
- for for fluid dynamics or computational fluid dynamics.
224
00:08:35.504 --> 00:08:37.365
- And and these, you know, these simulations
225
00:08:38.049 --> 00:08:38.549
- take
226
00:08:38.850 --> 00:08:41.170
- the the the flow that governs their dynamics,
227
00:08:41.170 --> 00:08:42.950
- the Navios Stokes equations generally,
228
00:08:43.330 --> 00:08:45.110
- and and applies them to
229
00:08:45.490 --> 00:08:46.149
- a discretized
230
00:08:46.769 --> 00:08:49.809
- car and and environment so that a computer
231
00:08:49.809 --> 00:08:50.950
- can solve that numerically.
232
00:08:51.394 --> 00:08:53.075
- And and as you suggest, you know, that
233
00:08:53.075 --> 00:08:54.855
- that that takes a lot of compute horsepower.
234
00:08:55.955 --> 00:08:58.375
- You know, the the the simulations can run,
235
00:08:59.235 --> 00:09:02.615
- if if if you're talking about, unsteady simulations,
236
00:09:02.835 --> 00:09:04.934
- they they they can run for days.
237
00:09:06.199 --> 00:09:08.039
- And and, you know, to get even to
238
00:09:08.039 --> 00:09:09.879
- get to those simulations, you obviously need to
239
00:09:09.879 --> 00:09:11.559
- draw a car and that's likely to be
240
00:09:11.559 --> 00:09:13.879
- in CAD and take some time. But these
241
00:09:13.879 --> 00:09:16.679
- numerical simulations are are pretty fickle. You know,
242
00:09:16.679 --> 00:09:18.940
- to get those numerics to work,
243
00:09:19.485 --> 00:09:21.725
- you need a perfectly closed volume to to
244
00:09:21.725 --> 00:09:24.285
- to solve the calculations, and that actually often
245
00:09:24.285 --> 00:09:26.365
- takes multiple days of work to go from
246
00:09:26.365 --> 00:09:28.205
- something that looks, you know, like a kind
247
00:09:28.205 --> 00:09:30.605
- of finished car on the screen to something
248
00:09:30.605 --> 00:09:32.365
- that you can get through a simulation. And
249
00:09:32.365 --> 00:09:34.049
- and and there's a hidden effort there, which
250
00:09:34.209 --> 00:09:34.709
- aerodynamic
251
00:09:35.089 --> 00:09:35.589
- aerodynamicists
252
00:09:35.889 --> 00:09:37.909
- like myself know only too well.
253
00:09:38.769 --> 00:09:41.110
- So so when we start talking about these
254
00:09:41.250 --> 00:09:43.169
- AI tools, you know, an important part of
255
00:09:43.169 --> 00:09:45.649
- speed up is is also in the the
256
00:09:45.649 --> 00:09:46.789
- the kind of the preparation
257
00:09:48.095 --> 00:09:49.794
- effort that goes into to
258
00:09:50.254 --> 00:09:51.315
- a a current simulation.
259
00:09:51.774 --> 00:09:53.774
- There there's also the post processing side. So,
260
00:09:53.774 --> 00:09:55.554
- you know, once you run that that simulation,
261
00:09:56.735 --> 00:09:58.735
- you'll you'll want to extract the the things
262
00:09:58.735 --> 00:10:00.095
- you really care about. It might be the
263
00:10:00.095 --> 00:10:01.394
- lift. It might be the downforce.
264
00:10:02.610 --> 00:10:04.450
- So so so that you generally set up
265
00:10:04.450 --> 00:10:06.309
- some kind of process around that.
266
00:10:07.250 --> 00:10:09.730
- And and also maybe to understand the flow
267
00:10:09.730 --> 00:10:12.230
- field, you you might take a few slices,
268
00:10:13.009 --> 00:10:14.595
- through through the airflow,
269
00:10:15.134 --> 00:10:16.835
- which you'd sort of save down,
270
00:10:17.615 --> 00:10:20.115
- because that that simulation file is normally
271
00:10:20.495 --> 00:10:21.315
- pretty huge,
272
00:10:22.095 --> 00:10:23.634
- and impractical to save.
273
00:10:24.735 --> 00:10:26.434
- And you said these are big calculations.
274
00:10:26.815 --> 00:10:29.100
- Are can can you do them in house?
275
00:10:29.659 --> 00:10:32.220
- Or do you have to rent time on
276
00:10:32.220 --> 00:10:33.039
- a supercomputer?
277
00:10:33.500 --> 00:10:35.679
- Or do you do a bit of both?
278
00:10:37.179 --> 00:10:38.159
- Yeah. I mean,
279
00:10:39.500 --> 00:10:40.000
- most,
280
00:10:40.860 --> 00:10:43.259
- kind of most automotive players will will will
281
00:10:43.259 --> 00:10:45.495
- have their HPC resource.
282
00:10:46.434 --> 00:10:48.674
- Oft often that is in house, and, you
283
00:10:48.674 --> 00:10:51.575
- know, these are these are big big investments,
284
00:10:52.595 --> 00:10:53.654
- but but increasingly,
285
00:10:54.355 --> 00:10:55.575
- kind of cloud provision,
286
00:10:56.274 --> 00:10:59.095
- and and as so it's moving towards GPUs,
287
00:11:00.160 --> 00:11:01.779
- on on GPU clusters,
288
00:11:04.080 --> 00:11:04.899
- are pretty typical.
289
00:11:07.440 --> 00:11:08.420
- And what about,
290
00:11:09.279 --> 00:11:11.220
- you know, so I I suppose proving
291
00:11:11.679 --> 00:11:13.860
- the the the worth of these simulations.
292
00:11:14.544 --> 00:11:16.384
- How do you do do you have to
293
00:11:16.384 --> 00:11:16.884
- compare
294
00:11:17.665 --> 00:11:19.524
- experimental data and,
295
00:11:20.625 --> 00:11:21.125
- simulations
296
00:11:21.504 --> 00:11:24.245
- in order to to know that you're on
297
00:11:24.465 --> 00:11:25.825
- the right track? I mean, do do you
298
00:11:25.825 --> 00:11:27.285
- have to actually build a model
299
00:11:27.710 --> 00:11:29.309
- of the car and and stick it in
300
00:11:29.309 --> 00:11:31.309
- a wind tunnel? I'm guessing that's probably what
301
00:11:31.309 --> 00:11:33.389
- you're trying to avoid because that would be
302
00:11:33.389 --> 00:11:36.610
- very expensive. So how how do you confirm
303
00:11:36.670 --> 00:11:38.610
- that your that your simulations
304
00:11:38.910 --> 00:11:39.730
- are working?
305
00:11:41.470 --> 00:11:41.970
- Yeah.
306
00:11:42.855 --> 00:11:43.355
- Absolutely,
307
00:11:44.534 --> 00:11:45.034
- experiments
308
00:11:45.975 --> 00:11:48.134
- are an important part of the picture. And,
309
00:11:48.534 --> 00:11:51.174
- if I kind of think about my formula
310
00:11:51.174 --> 00:11:51.835
- one days,
311
00:11:53.174 --> 00:11:55.495
- then, you know, we we we were using
312
00:11:55.495 --> 00:11:56.394
- the wind tunnel
313
00:11:57.549 --> 00:12:01.649
- very heavily to kind of correlate and tune
314
00:12:02.110 --> 00:12:03.629
- our our CFD. In the end, you know,
315
00:12:03.629 --> 00:12:05.570
- the CFD is a is a model of
316
00:12:05.710 --> 00:12:07.089
- a a model of your reality.
317
00:12:07.950 --> 00:12:10.370
- And and so, you know, having that experimental
318
00:12:10.509 --> 00:12:11.009
- data,
319
00:12:11.629 --> 00:12:14.004
- albeit it's, you know, it's it's time consuming
320
00:12:14.004 --> 00:12:15.305
- and it's and it's expensive
321
00:12:15.764 --> 00:12:17.945
- is is an important part of validation
322
00:12:18.245 --> 00:12:18.985
- for sure.
323
00:12:20.805 --> 00:12:22.325
- I see. But,
324
00:12:22.725 --> 00:12:24.245
- and and and so now I mean I
325
00:12:24.245 --> 00:12:26.325
- mean, maybe this is a good time to
326
00:12:26.325 --> 00:12:27.785
- to start talking about
327
00:12:28.320 --> 00:12:28.820
- AI.
328
00:12:29.279 --> 00:12:32.000
- So, David, before the advent of AI, how
329
00:12:32.000 --> 00:12:32.899
- were experimental
330
00:12:33.360 --> 00:12:34.180
- and simulated
331
00:12:34.639 --> 00:12:35.139
- aerodynamics
332
00:12:35.680 --> 00:12:38.259
- data used to design cars?
333
00:12:39.920 --> 00:12:41.300
- Great question. Thanks, Hamish.
334
00:12:42.345 --> 00:12:44.985
- So during my f one experience, although I
335
00:12:44.985 --> 00:12:47.465
- think the answer is similar for for sort
336
00:12:47.465 --> 00:12:49.004
- of passenger cars as well.
337
00:12:49.384 --> 00:12:51.384
- You know, when I started at Benetton in
338
00:12:51.384 --> 00:12:53.945
- 2000, it was really all about wind tunnel
339
00:12:53.945 --> 00:12:54.445
- testing,
340
00:12:55.350 --> 00:12:57.910
- and just doing that as much as possible
341
00:12:57.910 --> 00:12:59.210
- and as fast as possible.
342
00:12:59.590 --> 00:13:02.730
- But it was expensive. And, you know, even
343
00:13:03.029 --> 00:13:05.029
- with, you know, f one speed, it was
344
00:13:05.029 --> 00:13:07.450
- relatively slow weeks between loops
345
00:13:07.995 --> 00:13:10.235
- and limited in terms of the feedback that
346
00:13:10.235 --> 00:13:10.975
- you could get,
347
00:13:11.434 --> 00:13:13.835
- from those tests. You you'd get force from
348
00:13:13.835 --> 00:13:16.014
- the balance and maybe a few pressure sensors.
349
00:13:17.514 --> 00:13:20.315
- Then on the CFD side, we had, at
350
00:13:20.315 --> 00:13:22.659
- that stage, just one engineer, and their entire
351
00:13:22.659 --> 00:13:24.899
- focus actually was on the the airbox, which
352
00:13:24.899 --> 00:13:26.659
- is is the sort of feed behind the
353
00:13:26.659 --> 00:13:28.440
- driver's head going to the engine.
354
00:13:28.820 --> 00:13:30.899
- Otherwise, it was all wind tunnel. And there
355
00:13:30.899 --> 00:13:32.659
- was a lot of skepticism at the time
356
00:13:32.659 --> 00:13:34.600
- about that simulation with engineers
357
00:13:35.125 --> 00:13:37.225
- saying they didn't believe the CFD and and
358
00:13:37.605 --> 00:13:38.585
- not using it.
359
00:13:38.965 --> 00:13:40.904
- And then through the two thousands,
360
00:13:41.845 --> 00:13:43.445
- you know, we we saw that the the
361
00:13:43.445 --> 00:13:45.524
- teams that really thrived were the ones that
362
00:13:45.524 --> 00:13:47.465
- saw CFD as something complimentary.
363
00:13:47.909 --> 00:13:50.069
- You know, it was much, much faster than
364
00:13:50.069 --> 00:13:52.629
- than wind tunnel testing. It provided flow field
365
00:13:52.629 --> 00:13:54.470
- insight that that you couldn't get from the
366
00:13:54.470 --> 00:13:55.209
- wind tunnel.
367
00:13:55.669 --> 00:13:57.669
- So with results in a day rather than
368
00:13:57.669 --> 00:14:00.870
- weeks, that massively, you know, accelerated how you
369
00:14:00.870 --> 00:14:03.085
- could develop the car, and that insight that
370
00:14:03.085 --> 00:14:06.384
- you got really helped inform next steps. So
371
00:14:06.925 --> 00:14:08.845
- then if if we roll through to sort
372
00:14:08.845 --> 00:14:09.825
- of 2020,
373
00:14:11.245 --> 00:14:11.745
- then
374
00:14:12.125 --> 00:14:14.225
- your CFD was absolutely a given.
375
00:14:14.580 --> 00:14:15.480
- And and and, actually,
376
00:14:16.820 --> 00:14:19.059
- the the majority of team principals in f
377
00:14:19.059 --> 00:14:21.320
- one voted in an f one meeting
378
00:14:21.700 --> 00:14:24.019
- to turn off wind tunnels completely for for
379
00:14:24.019 --> 00:14:26.500
- cost saving. But, actually, I think that that
380
00:14:26.500 --> 00:14:28.740
- kinda misses the point, to one of your
381
00:14:28.740 --> 00:14:29.720
- previous questions.
382
00:14:30.054 --> 00:14:32.134
- You know, the these these tools are are
383
00:14:32.134 --> 00:14:34.695
- best when you use them together. And I
384
00:14:34.695 --> 00:14:37.095
- think we're living through the same moment with
385
00:14:37.095 --> 00:14:39.274
- with AI. You know, there's absolutely,
386
00:14:40.054 --> 00:14:40.795
- some skepticism,
387
00:14:42.615 --> 00:14:45.070
- and and and some some engineers see this
388
00:14:45.070 --> 00:14:46.990
- as a swap in for CFD. But I
389
00:14:46.990 --> 00:14:49.870
- think the most successful users I see are
390
00:14:49.870 --> 00:14:52.429
- are those that are are using AI as
391
00:14:52.429 --> 00:14:53.250
- another layer,
392
00:14:53.950 --> 00:14:56.750
- giving that massive acceleration that it offers, but
393
00:14:56.750 --> 00:14:59.230
- working alongside the the tools and expertise already
394
00:14:59.230 --> 00:14:59.715
- in place.
395
00:15:01.715 --> 00:15:04.434
- I see. And and may I suppose moving
396
00:15:04.434 --> 00:15:06.375
- on to AI, PhysicsX develops
397
00:15:07.394 --> 00:15:09.495
- large physics models for aerodynamics
398
00:15:10.115 --> 00:15:11.335
- and other applications.
399
00:15:11.809 --> 00:15:14.610
- What is a a large physics model? And
400
00:15:14.610 --> 00:15:16.309
- how does it differ from
401
00:15:16.690 --> 00:15:17.190
- conventional
402
00:15:17.649 --> 00:15:19.669
- computational approaches to aerodynamics?
403
00:15:21.490 --> 00:15:23.169
- Yeah. So so if I I kind of
404
00:15:23.169 --> 00:15:24.389
- start with the differences.
405
00:15:25.834 --> 00:15:26.334
- So
406
00:15:26.634 --> 00:15:29.274
- machine learning physics model's different in a in
407
00:15:29.274 --> 00:15:30.575
- a couple of different ways.
408
00:15:31.115 --> 00:15:32.654
- Unlike traditional computational
409
00:15:33.434 --> 00:15:33.934
- approaches,
410
00:15:34.315 --> 00:15:35.454
- machine learning model,
411
00:15:35.834 --> 00:15:37.774
- it it doesn't need to be told explicitly
412
00:15:37.914 --> 00:15:39.120
- what the laws of physics
413
00:15:39.600 --> 00:15:40.100
- are.
414
00:15:40.639 --> 00:15:42.879
- It learns it learns that from the data
415
00:15:42.879 --> 00:15:44.799
- it sees and and, you know, we we
416
00:15:44.799 --> 00:15:47.519
- use architectures that are suited to specific physics
417
00:15:47.519 --> 00:15:48.019
- problems.
418
00:15:48.559 --> 00:15:50.320
- And and and in doing that, it becomes
419
00:15:50.320 --> 00:15:53.174
- much more flexible than a traditional CFD tool.
420
00:15:53.975 --> 00:15:56.134
- It it doesn't require the same geometry preparation
421
00:15:56.134 --> 00:15:57.274
- we talked about earlier,
422
00:15:57.654 --> 00:15:59.654
- and and it can learn not just from
423
00:15:59.654 --> 00:16:02.534
- simulation data, but also from experimental data. So
424
00:16:02.534 --> 00:16:04.235
- it can actually bridge the gap.
425
00:16:06.160 --> 00:16:07.379
- It it can bridge the
426
00:16:07.840 --> 00:16:10.399
- gap, from from simulation world to kind of
427
00:16:10.399 --> 00:16:11.139
- ground truth.
428
00:16:12.160 --> 00:16:14.399
- It it it's got some obvious benefits in
429
00:16:14.399 --> 00:16:16.960
- terms of speed. Obviously, that that that's kind
430
00:16:16.960 --> 00:16:19.040
- of well known, and and and those speeds
431
00:16:19.040 --> 00:16:21.315
- can be up to, you know, 10,000 or
432
00:16:21.315 --> 00:16:23.735
- more times. So so really game changing.
433
00:16:24.434 --> 00:16:26.115
- And and that allows you to, as I
434
00:16:26.115 --> 00:16:26.615
- mentioned
435
00:16:27.235 --> 00:16:29.875
- in the example before, that really transforms what
436
00:16:29.875 --> 00:16:30.615
- you can do,
437
00:16:31.394 --> 00:16:34.274
- particularly when you get into solving multi physics
438
00:16:34.274 --> 00:16:35.334
- systems. So,
439
00:16:35.909 --> 00:16:36.970
- in in a traditional,
440
00:16:38.070 --> 00:16:39.909
- maybe in an OEM, you you you might
441
00:16:39.909 --> 00:16:41.049
- be solving problems
442
00:16:41.509 --> 00:16:43.529
- that that span across aerodynamics.
443
00:16:44.629 --> 00:16:46.309
- You've made it be in an engine design
444
00:16:46.309 --> 00:16:49.029
- where you've got aerodynamics involved, but you've also
445
00:16:49.029 --> 00:16:50.254
- got structures involved.
446
00:16:50.875 --> 00:16:53.514
- And and solving those different physics goes across
447
00:16:53.514 --> 00:16:55.934
- many groups and many groups of experts.
448
00:16:56.634 --> 00:16:58.095
- But with with these models,
449
00:16:59.035 --> 00:17:01.434
- then then you can solve those physics all
450
00:17:01.434 --> 00:17:03.660
- at the same time. And because the models
451
00:17:03.660 --> 00:17:05.440
- are more accessible than,
452
00:17:05.980 --> 00:17:07.599
- than than sort of simulation
453
00:17:08.140 --> 00:17:08.640
- tools,
454
00:17:08.940 --> 00:17:10.380
- then you can put them direct in the
455
00:17:10.380 --> 00:17:11.339
- hands of a designer.
456
00:17:11.660 --> 00:17:13.900
- And and that really starts to, again, change
457
00:17:13.900 --> 00:17:16.140
- the way you're working rather than just doing
458
00:17:16.140 --> 00:17:18.365
- things a bit faster. You you really start
459
00:17:18.365 --> 00:17:20.224
- to rethink how you're working.
460
00:17:20.764 --> 00:17:22.845
- And and then finally, another aspect of these
461
00:17:22.845 --> 00:17:24.144
- models is,
462
00:17:24.764 --> 00:17:26.845
- that that they've they've got this concept of
463
00:17:26.845 --> 00:17:27.345
- uncertainty
464
00:17:27.724 --> 00:17:28.224
- quantification.
465
00:17:28.605 --> 00:17:31.105
- So they they know how confident they are,
466
00:17:32.069 --> 00:17:32.569
- and
467
00:17:32.869 --> 00:17:34.549
- and and that's not just important from an
468
00:17:34.549 --> 00:17:36.390
- engineer's point of view because we need to
469
00:17:36.390 --> 00:17:37.450
- be building trust
470
00:17:37.750 --> 00:17:38.490
- in in,
471
00:17:39.029 --> 00:17:40.549
- in in the kind of use of of
472
00:17:40.549 --> 00:17:41.210
- the results.
473
00:17:42.230 --> 00:17:44.549
- But also from from the model point of
474
00:17:44.549 --> 00:17:46.890
- view, we can use that uncertainty to drive
475
00:17:47.204 --> 00:17:49.684
- the the simulations that the model needs to
476
00:17:49.684 --> 00:17:51.605
- be accurate in the places that it needs
477
00:17:51.605 --> 00:17:53.125
- to be. So we can have these, what
478
00:17:53.125 --> 00:17:54.664
- we call, active learning loops,
479
00:17:55.444 --> 00:17:58.585
- that that automatically trigger new set new simulations
480
00:17:58.724 --> 00:18:00.505
- when when a model becomes unsure.
481
00:18:03.279 --> 00:18:04.559
- And David, you,
482
00:18:04.880 --> 00:18:06.579
- and your colleagues at PhysicsX
483
00:18:06.880 --> 00:18:10.819
- have released a study called scaling physics AI
484
00:18:11.119 --> 00:18:11.940
- for automotive
485
00:18:12.319 --> 00:18:12.819
- aerodynamics.
486
00:18:13.759 --> 00:18:16.579
- And it argues that publicly available datasets
487
00:18:17.054 --> 00:18:17.875
- of automotive
488
00:18:18.254 --> 00:18:18.754
- simulations
489
00:18:19.534 --> 00:18:21.554
- are not sufficient to train
490
00:18:21.934 --> 00:18:23.474
- large physics models.
491
00:18:24.174 --> 00:18:26.274
- Why is this? And what are the consequences
492
00:18:26.494 --> 00:18:27.634
- for those using
493
00:18:28.095 --> 00:18:29.875
- AI to design cars?
494
00:18:31.390 --> 00:18:33.869
- Yeah. So so I should maybe give one
495
00:18:33.869 --> 00:18:36.589
- last piece of detail about the the these
496
00:18:36.589 --> 00:18:38.829
- types of models and and just to sort
497
00:18:38.829 --> 00:18:41.789
- of transition to the automotive large physics model
498
00:18:41.789 --> 00:18:42.450
- that we've,
499
00:18:43.069 --> 00:18:44.990
- developed at PhysicsX because I think that gives
500
00:18:44.990 --> 00:18:45.730
- some context.
501
00:18:46.855 --> 00:18:49.414
- And, you know, this is a pretrained model,
502
00:18:49.414 --> 00:18:51.335
- so it's already seen a huge amount of
503
00:18:51.335 --> 00:18:53.335
- data. And and that means it's something that
504
00:18:53.335 --> 00:18:55.335
- we can take to a customer that's that's
505
00:18:55.335 --> 00:18:56.535
- sort of ready to go out of the
506
00:18:56.535 --> 00:18:58.714
- box and and can predict aerodynamics
507
00:18:59.130 --> 00:19:00.829
- of previously unseen vehicles.
508
00:19:02.170 --> 00:19:04.329
- And and and a a really powerful aspect
509
00:19:04.329 --> 00:19:06.570
- of that is when we start fine tuning
510
00:19:06.570 --> 00:19:07.930
- that model. So we take the model that
511
00:19:07.930 --> 00:19:09.849
- we've developed in the house. We we fine
512
00:19:09.849 --> 00:19:12.944
- tune it for a given customer's needs or,
513
00:19:13.585 --> 00:19:17.044
- specifics of their car type and simulation approach.
514
00:19:17.505 --> 00:19:20.464
- But underpinning all of that is is is
515
00:19:20.464 --> 00:19:22.784
- the data, the high quality data that we
516
00:19:22.784 --> 00:19:25.024
- need to build that model. And and that
517
00:19:25.024 --> 00:19:26.404
- really comes to your question
518
00:19:28.430 --> 00:19:28.930
- about
519
00:19:30.670 --> 00:19:31.809
- the the need for,
520
00:19:33.150 --> 00:19:34.930
- the the the need for a dataset,
521
00:19:35.950 --> 00:19:36.450
- and
522
00:19:36.830 --> 00:19:37.330
- our,
523
00:19:39.150 --> 00:19:41.070
- our own our own dataset. So there's a
524
00:19:41.070 --> 00:19:42.934
- there's a couple of well publicized,
525
00:19:44.115 --> 00:19:47.394
- publicly available datasets, driver net and luminary shift
526
00:19:47.394 --> 00:19:47.894
- SUV.
527
00:19:49.075 --> 00:19:50.454
- And those were,
528
00:19:51.154 --> 00:19:53.335
- both derived from sort of parametric
529
00:19:53.794 --> 00:19:54.294
- morphing
530
00:19:54.994 --> 00:19:57.255
- of of a small number of base configurations.
531
00:19:58.180 --> 00:19:59.940
- And that that morphing is is a kind
532
00:19:59.940 --> 00:20:02.019
- of where you take a geometric mesh and
533
00:20:02.019 --> 00:20:03.400
- you kind of push and pull,
534
00:20:03.859 --> 00:20:05.640
- to to give geometric variation.
535
00:20:07.140 --> 00:20:08.840
- And, you know, these,
536
00:20:10.580 --> 00:20:11.960
- the these these variations,
537
00:20:13.195 --> 00:20:16.015
- obviously give give some diversity, but what we,
538
00:20:16.075 --> 00:20:17.515
- you know, what we saw when we we
539
00:20:17.515 --> 00:20:19.055
- started digging into these models
540
00:20:19.434 --> 00:20:22.075
- was that they performed well on on when
541
00:20:22.075 --> 00:20:24.875
- we tested them on geometries, which were very
542
00:20:24.875 --> 00:20:26.015
- similar to those,
543
00:20:26.795 --> 00:20:27.295
- datasets.
544
00:20:27.929 --> 00:20:29.769
- But but really, they started to fail quite
545
00:20:29.769 --> 00:20:31.869
- badly when we looked at realistic
546
00:20:32.569 --> 00:20:35.210
- customer geometries, which which were of a different
547
00:20:35.210 --> 00:20:35.710
- design.
548
00:20:38.730 --> 00:20:41.794
- So to to address this shortcoming that that
549
00:20:41.794 --> 00:20:43.815
- you've identified in your study,
550
00:20:44.115 --> 00:20:47.734
- PhysicsX has created the p x NETCAR
551
00:20:48.355 --> 00:20:48.855
- proprietary
552
00:20:49.315 --> 00:20:49.815
- dataset.
553
00:20:50.835 --> 00:20:52.914
- How did you develop this, and how does
554
00:20:52.914 --> 00:20:54.775
- it compare to public datasets?
555
00:20:56.670 --> 00:20:57.390
- Yeah. So,
556
00:20:57.950 --> 00:21:00.269
- we we kind of having having drawn that
557
00:21:00.269 --> 00:21:02.690
- conclusion that that, you know, we we needed
558
00:21:02.910 --> 00:21:03.410
- this
559
00:21:04.029 --> 00:21:05.809
- more diversity, so we built,
560
00:21:06.109 --> 00:21:07.009
- what we call,
561
00:21:07.470 --> 00:21:09.934
- the the Physics x Data Factory, and that's
562
00:21:09.934 --> 00:21:11.315
- really an automated pipeline
563
00:21:11.855 --> 00:21:13.474
- that can run many, many,
564
00:21:14.255 --> 00:21:14.755
- simulations.
565
00:21:15.134 --> 00:21:18.035
- In this case, STAR CCM plus runs,
566
00:21:18.414 --> 00:21:21.075
- Reynolds' average average Navios Stokes simulations.
567
00:21:22.589 --> 00:21:24.349
- So, you know, we we got a hold
568
00:21:24.349 --> 00:21:25.730
- of 450
569
00:21:25.789 --> 00:21:27.410
- quite distinct car shapes,
570
00:21:28.109 --> 00:21:30.450
- from large scale three d model libraries.
571
00:21:31.230 --> 00:21:32.289
- And then we
572
00:21:32.910 --> 00:21:35.044
- produced our own kinda algorithmic
573
00:21:35.424 --> 00:21:37.605
- morphing approach so that we could create
574
00:21:38.065 --> 00:21:40.384
- around a 100 variants for for each of
575
00:21:40.384 --> 00:21:43.204
- these baselines using a a tool called Blender,
576
00:21:43.825 --> 00:21:45.825
- which is an open source three d modeling
577
00:21:45.825 --> 00:21:46.325
- tool.
578
00:21:47.009 --> 00:21:49.109
- And and we used a Latin hypertube
579
00:21:49.730 --> 00:21:51.970
- sampling approach so that we knew we were
580
00:21:51.970 --> 00:21:54.150
- kind of creating a a statistically,
581
00:21:55.890 --> 00:21:57.990
- good coverage of of these geometries.
582
00:21:59.105 --> 00:22:01.845
- And then we we rigorously validated,
583
00:22:02.705 --> 00:22:04.144
- those results in terms of the kind of
584
00:22:04.144 --> 00:22:06.945
- convergence and the correctness of the physics knowing
585
00:22:06.945 --> 00:22:07.924
- that, you know,
586
00:22:08.384 --> 00:22:10.484
- the the the quality of that data was
587
00:22:10.545 --> 00:22:12.865
- critical to the performance of the model at
588
00:22:12.865 --> 00:22:13.525
- the end.
589
00:22:13.970 --> 00:22:16.710
- So that that allowed us to build PXnetcar,
590
00:22:17.809 --> 00:22:18.869
- which now comprises
591
00:22:19.170 --> 00:22:20.869
- over 30,000 simulations
592
00:22:21.410 --> 00:22:23.250
- across those 450
593
00:22:23.250 --> 00:22:23.750
- distinct,
594
00:22:24.609 --> 00:22:26.434
- geometries. And, you know,
595
00:22:26.894 --> 00:22:29.554
- comparing back to the driver net and luminary
596
00:22:30.014 --> 00:22:31.774
- shift, which is where we started with, you
597
00:22:31.774 --> 00:22:32.355
- know, between
598
00:22:33.214 --> 00:22:35.714
- those two, that they they gave us 10,
599
00:22:36.494 --> 00:22:39.375
- kind of 10 distinct geometry start points. So
600
00:22:39.375 --> 00:22:41.554
- it's it's really a complete game changer
601
00:22:41.910 --> 00:22:43.289
- scaling up that much.
602
00:22:43.990 --> 00:22:45.830
- And and, you know, we wanted to be
603
00:22:45.830 --> 00:22:46.890
- quite quantitative
604
00:22:47.269 --> 00:22:50.390
- about the, the diversity we were generating. So
605
00:22:50.390 --> 00:22:52.730
- we use something called pairwise chamfer distance,
606
00:22:53.509 --> 00:22:56.170
- which basically measures between different geometries
607
00:22:56.605 --> 00:22:58.284
- the the the kind of distance between all
608
00:22:58.284 --> 00:23:00.365
- the points, and and it gives a a
609
00:23:00.365 --> 00:23:01.184
- way of quantifying
610
00:23:01.565 --> 00:23:04.365
- that diversity. And we could see as we
611
00:23:04.365 --> 00:23:06.444
- build up that dataset that we were,
612
00:23:07.325 --> 00:23:09.325
- that we were really building a a very
613
00:23:09.325 --> 00:23:09.825
- diverse,
614
00:23:10.970 --> 00:23:12.809
- set of classes, not just sort of minor
615
00:23:12.809 --> 00:23:13.309
- morphs.
616
00:23:14.009 --> 00:23:16.269
- And then when we started applying that,
617
00:23:16.890 --> 00:23:19.529
- a large physics model in in practice, then
618
00:23:19.529 --> 00:23:22.750
- that very clearly translated into much greater generalization
619
00:23:23.210 --> 00:23:25.115
- performance. So, you you know, when we started
620
00:23:25.174 --> 00:23:27.355
- showing it new geometries, new cars,
621
00:23:28.214 --> 00:23:31.015
- that then the performance was much much higher
622
00:23:31.015 --> 00:23:33.095
- than than than the models that we'd,
623
00:23:33.494 --> 00:23:35.355
- trained on those narrower, luminary,
624
00:23:36.375 --> 00:23:38.315
- and, and and driver net datasets.
625
00:23:40.410 --> 00:23:42.970
- And and we've been talking about cars, David,
626
00:23:42.970 --> 00:23:46.349
- but PhysicsX has a range of interests outside
627
00:23:46.410 --> 00:23:49.130
- of vehicle design. Can you talk about some
628
00:23:49.130 --> 00:23:51.390
- of the other sectors that you're developing
629
00:23:52.154 --> 00:23:53.215
- AI tools for?
630
00:23:54.075 --> 00:23:55.355
- Yeah. Absolutely. Yeah.
631
00:23:55.914 --> 00:23:58.075
- I mean, we work on some incredibly exciting
632
00:23:58.075 --> 00:23:58.575
- projects.
633
00:23:59.195 --> 00:24:00.955
- And, you know, many of these we can't
634
00:24:00.955 --> 00:24:03.055
- talk about because they're they're customer projects.
635
00:24:03.355 --> 00:24:05.035
- I happen to be working on one which
636
00:24:05.035 --> 00:24:07.579
- is, has been publicly announced, so that that's
637
00:24:07.579 --> 00:24:08.319
- quite helpful.
638
00:24:08.859 --> 00:24:10.940
- But starting more generally, you know, our our
639
00:24:10.940 --> 00:24:13.519
- models add most value in in in sectors
640
00:24:13.579 --> 00:24:14.079
- where
641
00:24:14.380 --> 00:24:15.679
- complex physics simulation
642
00:24:15.980 --> 00:24:18.460
- is right at the heart of engineering decision
643
00:24:18.460 --> 00:24:20.914
- making or operation. And, generally, you know, the
644
00:24:20.914 --> 00:24:22.595
- time that we spoke about earlier to solve
645
00:24:22.595 --> 00:24:23.335
- those simulations
646
00:24:23.795 --> 00:24:25.974
- can be a real bottleneck for for engineers
647
00:24:26.035 --> 00:24:27.174
- or product development.
648
00:24:28.115 --> 00:24:30.615
- Commonly, we see this across aerospace and semiconductors,
649
00:24:31.154 --> 00:24:33.240
- energy, and materials. So, you know, this is
650
00:24:33.240 --> 00:24:35.099
- where we we're doing a lot of work.
651
00:24:35.799 --> 00:24:38.119
- And and, also, it's not just about design,
652
00:24:38.119 --> 00:24:40.119
- but because these models can solve more or
653
00:24:40.119 --> 00:24:43.339
- less real time, they become ideal for manufacturing
654
00:24:43.720 --> 00:24:45.255
- and and operation processes.
655
00:24:46.134 --> 00:24:46.634
- You
656
00:24:47.015 --> 00:24:48.615
- know, just to give a small sense of
657
00:24:48.615 --> 00:24:50.394
- the complexity, a recent project
658
00:24:50.855 --> 00:24:53.494
- we developed was a a plasma reactor for
659
00:24:53.494 --> 00:24:57.515
- for converting industrial waste gases into synthetic fuels.
660
00:24:57.619 --> 00:25:00.340
- So basically taking, gases including c o two
661
00:25:00.340 --> 00:25:02.680
- and then converting that into carbon monoxide,
662
00:25:03.779 --> 00:25:05.779
- which which then went on into these,
663
00:25:07.299 --> 00:25:08.519
- synthetic hydrocarbons.
664
00:25:09.924 --> 00:25:11.605
- And, you know, we had to model all
665
00:25:11.605 --> 00:25:14.085
- the chemical reactions, the mixing of flow through
666
00:25:14.085 --> 00:25:14.744
- the reactor,
667
00:25:15.684 --> 00:25:17.845
- and and the thermal modeling, of course, to
668
00:25:17.845 --> 00:25:19.924
- to make sure that kind of critical components
669
00:25:19.924 --> 00:25:20.664
- didn't overheat.
670
00:25:21.285 --> 00:25:23.045
- And with with this technology, we can do
671
00:25:23.045 --> 00:25:24.025
- all of that simultaneously.
672
00:25:24.565 --> 00:25:25.140
- So it's
673
00:25:25.619 --> 00:25:27.160
- massive, massive speed up.
674
00:25:27.779 --> 00:25:29.940
- And, you know, we we we reduced the
675
00:25:29.940 --> 00:25:32.019
- r and d process by a couple of
676
00:25:32.019 --> 00:25:34.259
- years, and and in the end, had a
677
00:25:34.259 --> 00:25:36.820
- machine that was giving a tenfold increase in
678
00:25:36.820 --> 00:25:39.779
- react productivity, which is, you know, an absolute
679
00:25:39.779 --> 00:25:40.424
- game changer.
680
00:25:41.705 --> 00:25:43.545
- Moving moving on to the project that I'm
681
00:25:43.545 --> 00:25:44.365
- involved in,
682
00:25:45.785 --> 00:25:48.825
- so so this is working with GB one,
683
00:25:48.825 --> 00:25:51.305
- which is The UK's entrant to the twenty
684
00:25:51.305 --> 00:25:52.845
- seven America's Cup.
685
00:25:53.930 --> 00:25:55.930
- And it's I think it's an interesting example
686
00:25:55.930 --> 00:25:58.109
- because, again, it's it's a very advanced engineering
687
00:25:58.170 --> 00:26:00.809
- workflow. And and and for some context for
688
00:26:00.809 --> 00:26:02.029
- those maybe not aware,
689
00:26:02.650 --> 00:26:06.109
- America's cup boats are 75 foot hydrofoiling
690
00:26:06.490 --> 00:26:07.335
- sailing yachts,
691
00:26:08.054 --> 00:26:10.154
- often better described as flying machines.
692
00:26:10.454 --> 00:26:12.534
- You know, they they they go pretty crazy
693
00:26:12.534 --> 00:26:15.015
- speeds in excess of four and a half
694
00:26:15.015 --> 00:26:17.335
- times wind speed, which as an engineer takes
695
00:26:17.335 --> 00:26:19.034
- a bit of wrapping your head around,
696
00:26:19.849 --> 00:26:21.930
- which means they go they go above a
697
00:26:21.930 --> 00:26:23.470
- 100 kilometers an hour.
698
00:26:23.849 --> 00:26:25.450
- And I was lucky to join the team,
699
00:26:25.690 --> 00:26:27.369
- for some testing in Sardinia a couple of
700
00:26:27.369 --> 00:26:29.450
- weeks ago. And and then you really do
701
00:26:29.450 --> 00:26:31.049
- get a sense of of the speed when
702
00:26:31.049 --> 00:26:32.890
- this boat is is moving towards you at
703
00:26:32.890 --> 00:26:34.269
- that that kind of speed.
704
00:26:35.525 --> 00:26:37.525
- And and, you know, like in f one,
705
00:26:37.525 --> 00:26:40.244
- it's very much a competition between engine engineering
706
00:26:40.244 --> 00:26:40.744
- departments.
707
00:26:41.924 --> 00:26:44.904
- So they've got very, very refined development processes.
708
00:26:45.365 --> 00:26:47.444
- But one of the big remaining challenges is
709
00:26:47.444 --> 00:26:48.825
- just a massive dimensionality,
710
00:26:50.460 --> 00:26:52.539
- of of the design and operating space that
711
00:26:52.539 --> 00:26:54.700
- that the engineers have to contend with. And
712
00:26:54.700 --> 00:26:57.500
- this is where boats are different from f
713
00:26:57.500 --> 00:26:58.160
- one cars.
714
00:26:59.339 --> 00:27:00.940
- You know, f one cars, you've got the
715
00:27:00.940 --> 00:27:03.519
- steering and and accelerator brake pedal as input.
716
00:27:03.740 --> 00:27:06.945
- But with the boats, you've got multiple controls
717
00:27:06.945 --> 00:27:08.005
- across the foils,
718
00:27:08.384 --> 00:27:09.045
- the sails,
719
00:27:09.825 --> 00:27:12.065
- you know, the the sail shape itself, you've
720
00:27:12.065 --> 00:27:14.305
- got many controls, plus you've got the the
721
00:27:14.305 --> 00:27:15.045
- wind effect.
722
00:27:15.664 --> 00:27:17.605
- So not only are there all these design
723
00:27:18.160 --> 00:27:20.160
- parameters, but there's also all these operational ones
724
00:27:20.160 --> 00:27:22.240
- as well. And and that's kind of where
725
00:27:22.240 --> 00:27:22.640
- we've found,
726
00:27:24.319 --> 00:27:25.920
- you know, a a a a really good
727
00:27:25.920 --> 00:27:28.579
- fit between GP one and PhysicsX because
728
00:27:29.039 --> 00:27:29.539
- our,
729
00:27:30.079 --> 00:27:30.819
- kind of
730
00:27:31.119 --> 00:27:32.414
- physics aware architecture
731
00:27:32.894 --> 00:27:35.634
- can learn all those complex interactions, aerodynamics,
732
00:27:36.015 --> 00:27:36.515
- hydrodynamics,
733
00:27:38.494 --> 00:27:40.515
- and and predict near real time
734
00:27:40.894 --> 00:27:42.654
- both the flow around the car, but also
735
00:27:42.654 --> 00:27:43.154
- the,
736
00:27:43.775 --> 00:27:45.839
- the the the loads the loads on the
737
00:27:46.079 --> 00:27:48.019
- sorry. Around the boat, I should say.
738
00:27:48.799 --> 00:27:50.960
- And also the the the the loads that
739
00:27:50.960 --> 00:27:53.599
- that then feed into GB one simulation so
740
00:27:53.599 --> 00:27:55.200
- we can get a a a kind of
741
00:27:55.200 --> 00:27:56.259
- performance output.
742
00:27:57.200 --> 00:27:58.880
- And and and also in, you know, in
743
00:27:58.880 --> 00:28:00.615
- the structural domain as well. So
744
00:28:01.494 --> 00:28:03.355
- we we can give a really kinda complete
745
00:28:03.575 --> 00:28:04.075
- optimization.
746
00:28:05.974 --> 00:28:07.434
- So, you know, that that's,
747
00:28:07.974 --> 00:28:10.134
- using our platform, taking the learning from the
748
00:28:10.134 --> 00:28:12.454
- research group, the, you know, the the the
749
00:28:12.454 --> 00:28:12.954
- LPM
750
00:28:14.099 --> 00:28:16.200
- development that we spoke about, we can directly
751
00:28:16.259 --> 00:28:19.140
- translate very quickly into something that's been useful
752
00:28:19.140 --> 00:28:20.519
- for the GB one team.
753
00:28:21.859 --> 00:28:22.359
- And,
754
00:28:22.980 --> 00:28:24.579
- you know, in in in doing that, we
755
00:28:24.660 --> 00:28:27.299
- we're building very much on their expertise and
756
00:28:27.299 --> 00:28:28.679
- their physical insight.
757
00:28:29.555 --> 00:28:31.335
- But with with such a vast,
758
00:28:31.634 --> 00:28:33.414
- you know, vast space to explore,
759
00:28:33.875 --> 00:28:36.055
- the these tools are a complete game changer.
760
00:28:37.154 --> 00:28:37.555
- And,
761
00:28:38.275 --> 00:28:38.934
- you know,
762
00:28:39.875 --> 00:28:41.315
- one one of one of the points we
763
00:28:41.315 --> 00:28:43.789
- touched on earlier was, you know, bound the
764
00:28:43.789 --> 00:28:45.789
- the confidence of these models. And when we
765
00:28:45.789 --> 00:28:47.410
- want to start exploring such,
766
00:28:48.430 --> 00:28:49.170
- fast spaces,
767
00:28:49.549 --> 00:28:51.650
- we can, again, we can use that uncertainty
768
00:28:51.869 --> 00:28:54.349
- to kind of drive where we collect data.
769
00:28:54.349 --> 00:28:57.835
- So we we we're starting relatively small with
770
00:28:57.835 --> 00:28:58.494
- the dataset,
771
00:28:59.195 --> 00:29:01.835
- and and then as we start optimizing and
772
00:29:01.835 --> 00:29:04.555
- the optimizer is exploring the design space, we
773
00:29:04.555 --> 00:29:06.174
- can collect more and more data.
774
00:29:06.555 --> 00:29:06.955
- So,
775
00:29:07.355 --> 00:29:09.269
- you know, what we're seeing is is all
776
00:29:09.269 --> 00:29:10.549
- of the all of the,
777
00:29:11.190 --> 00:29:12.869
- lessons that we've learned in the,
778
00:29:13.590 --> 00:29:16.309
- large physics model are are giving us this
779
00:29:16.309 --> 00:29:18.549
- leg up with with our GPU one work,
780
00:29:18.549 --> 00:29:20.090
- and that's that's really exciting.
781
00:29:21.174 --> 00:29:22.934
- When when you're doing the,
782
00:29:23.974 --> 00:29:25.575
- the the work, you know, with the with
783
00:29:25.575 --> 00:29:26.234
- the boat,
784
00:29:26.615 --> 00:29:29.515
- are you are you just looking at
785
00:29:30.775 --> 00:29:31.275
- design,
786
00:29:31.575 --> 00:29:33.195
- or are you also
787
00:29:33.734 --> 00:29:34.529
- looking at
788
00:29:36.210 --> 00:29:38.609
- the the way in which the boat can
789
00:29:38.609 --> 00:29:39.349
- be sailed?
790
00:29:39.890 --> 00:29:41.190
- You know, looking for,
791
00:29:42.210 --> 00:29:44.210
- I don't know, maybe different ways of doing
792
00:29:44.210 --> 00:29:47.090
- things or the best way of of sailing
793
00:29:47.090 --> 00:29:48.630
- the boat in addition
794
00:29:49.154 --> 00:29:49.654
- to,
795
00:29:51.555 --> 00:29:52.695
- I suppose, adjust,
796
00:29:53.075 --> 00:29:55.075
- for lack of a better term, the the
797
00:29:55.075 --> 00:29:55.894
- fluid dynamics
798
00:29:56.515 --> 00:29:58.835
- of the boat? Or is that is is
799
00:29:58.835 --> 00:30:00.994
- that something different or maybe something that you
800
00:30:00.994 --> 00:30:02.660
- could look at in the future?
801
00:30:03.440 --> 00:30:05.759
- Well, yeah. Great question. And and maybe there's
802
00:30:05.759 --> 00:30:07.299
- a bit of both there because,
803
00:30:07.680 --> 00:30:09.440
- you know, even at the design stage, if
804
00:30:09.440 --> 00:30:11.279
- we wanna optimize the design, we need to
805
00:30:11.279 --> 00:30:13.220
- consider how the boat's gonna be sailed.
806
00:30:13.855 --> 00:30:15.134
- You know, we we can't do it in
807
00:30:15.134 --> 00:30:17.134
- isolation if we want to really kind of
808
00:30:17.134 --> 00:30:19.774
- get to a optimum performance. So, you know,
809
00:30:19.774 --> 00:30:20.755
- we are considering,
810
00:30:22.414 --> 00:30:25.394
- many aspects of how the boat will sail,
811
00:30:25.615 --> 00:30:27.375
- you know, as we're as we're going through
812
00:30:27.375 --> 00:30:28.434
- that design iteration.
813
00:30:30.869 --> 00:30:32.950
- But to your, you know, to your point,
814
00:30:32.950 --> 00:30:34.710
- we we can also use these types of
815
00:30:34.710 --> 00:30:37.049
- models in operations. So we're we're we're absolutely
816
00:30:37.190 --> 00:30:39.269
- thinking about, you know, how we might be
817
00:30:39.269 --> 00:30:41.684
- able to, kind of take the next step
818
00:30:41.684 --> 00:30:42.184
- and,
819
00:30:42.964 --> 00:30:45.144
- deploy these models, whether that's sort of,
820
00:30:46.565 --> 00:30:48.644
- give it giving kind of an almost advice
821
00:30:48.644 --> 00:30:50.664
- to the sailors on on on optimum,
822
00:30:51.285 --> 00:30:54.085
- sailing strategies or or or in in in
823
00:30:54.085 --> 00:30:54.585
- control.
824
00:30:57.200 --> 00:30:59.920
- And finally, here at Physics World, we're always
825
00:30:59.920 --> 00:31:01.619
- really interested in exploring
826
00:31:02.240 --> 00:31:02.740
- careers
827
00:31:03.200 --> 00:31:04.019
- for physicists.
828
00:31:04.720 --> 00:31:06.820
- And, I'm sure that some of our listeners
829
00:31:07.519 --> 00:31:09.914
- would be interested in a career in the
830
00:31:09.914 --> 00:31:12.815
- commercial development of large physics models.
831
00:31:13.355 --> 00:31:14.654
- What sort of employment
832
00:31:15.115 --> 00:31:17.134
- prospects are there in the sector?
833
00:31:17.595 --> 00:31:19.674
- And could could you maybe give a a
834
00:31:19.674 --> 00:31:22.875
- bit of career advice for someone who who'd
835
00:31:22.875 --> 00:31:24.495
- like to to get a job,
836
00:31:25.529 --> 00:31:26.750
- developing these models?
837
00:31:28.089 --> 00:31:31.450
- Yeah. Absolutely. And, you know, my own journey
838
00:31:31.450 --> 00:31:33.690
- to physics x is is is maybe not
839
00:31:33.690 --> 00:31:34.669
- a typical one.
840
00:31:36.329 --> 00:31:38.565
- But but maybe maybe there's a little bit,
841
00:31:38.644 --> 00:31:40.484
- you know, in in that journey that I
842
00:31:40.484 --> 00:31:42.105
- I still sort of take away.
843
00:31:42.804 --> 00:31:43.544
- You know,
844
00:31:44.004 --> 00:31:45.365
- when when I started out in f one,
845
00:31:45.365 --> 00:31:47.125
- it was definitely not a well trodden path.
846
00:31:47.125 --> 00:31:48.484
- And, you know, I was writing a lot
847
00:31:48.484 --> 00:31:50.484
- of letters, and I was looking for any
848
00:31:50.484 --> 00:31:51.539
- break I could get.
849
00:31:52.019 --> 00:31:53.859
- And and and, actually, my break came from
850
00:31:53.859 --> 00:31:56.180
- a rejection from what was then British American
851
00:31:56.180 --> 00:31:58.420
- Racing. You know, they wanted to employ some
852
00:31:58.420 --> 00:32:00.180
- of the experience, but they,
853
00:32:01.059 --> 00:32:03.059
- they they were on a notice period. So
854
00:32:03.059 --> 00:32:05.059
- there was, there there was a a a
855
00:32:05.059 --> 00:32:06.740
- few months where there was a vacant seat.
856
00:32:06.740 --> 00:32:07.515
- And so I
857
00:32:07.994 --> 00:32:10.075
- said, okay. Can I can I just come
858
00:32:10.075 --> 00:32:12.154
- and, you know, sit there and and learn
859
00:32:12.154 --> 00:32:13.914
- for a bit? And and and, actually, I
860
00:32:13.914 --> 00:32:15.674
- had the chance while I was there to
861
00:32:15.674 --> 00:32:18.015
- go and do some, testing in Italy.
862
00:32:18.795 --> 00:32:19.755
- We we were,
863
00:32:20.559 --> 00:32:23.599
- basically kind of putting wires across the the
864
00:32:23.599 --> 00:32:26.480
- race car to understand deflection effects. And then
865
00:32:26.480 --> 00:32:28.080
- I learned a huge amount. It was a
866
00:32:28.080 --> 00:32:30.240
- a great education. And, you know, it was
867
00:32:30.240 --> 00:32:32.755
- a a lesson definitely for me in, you
868
00:32:32.755 --> 00:32:35.154
- know, looking for those opportunities even when they
869
00:32:35.154 --> 00:32:36.855
- come in in unexpected ways.
870
00:32:38.034 --> 00:32:39.335
- And, you know, that
871
00:32:39.875 --> 00:32:41.875
- gave me my career in f one and
872
00:32:41.875 --> 00:32:43.335
- and then the the opportunity,
873
00:32:44.275 --> 00:32:47.559
- and the connection with PhysicsX also came with
874
00:32:47.559 --> 00:32:48.299
- some serendipity.
875
00:32:49.320 --> 00:32:51.740
- You know, it wasn't an opportunity I anticipated,
876
00:32:51.880 --> 00:32:53.559
- but, you know, after my experience as a
877
00:32:53.559 --> 00:32:55.980
- customer, not just the the impact
878
00:32:56.279 --> 00:32:57.559
- that I saw in f one, but the
879
00:32:57.559 --> 00:32:59.660
- scalability of this of a wider industry.
880
00:33:00.025 --> 00:33:01.464
- You know, I I knew it would be
881
00:33:01.464 --> 00:33:03.944
- an amazing place to be. And so, again,
882
00:33:03.944 --> 00:33:05.484
- kind of looking for that opportunity
883
00:33:06.025 --> 00:33:07.804
- to to kind of make the connections.
884
00:33:09.545 --> 00:33:10.744
- And now, you know, now I am at
885
00:33:10.744 --> 00:33:12.744
- PhysicsX, and I'm I'm seeing, you know, the
886
00:33:12.744 --> 00:33:14.744
- work that we're doing on on large physics
887
00:33:14.744 --> 00:33:15.244
- models.
888
00:33:15.730 --> 00:33:17.269
- You know, it's incredibly multidisciplinary.
889
00:33:19.009 --> 00:33:19.509
- Obviously,
890
00:33:19.890 --> 00:33:21.269
- it needs domain expertise.
891
00:33:22.769 --> 00:33:25.509
- Also, you know, we we need simulation engineering
892
00:33:25.569 --> 00:33:27.970
- to generate and validate our our datasets. We
893
00:33:27.970 --> 00:33:29.875
- talked about how important they are. And the
894
00:33:29.875 --> 00:33:33.394
- machine learning research spanning areas like geometric learning,
895
00:33:33.394 --> 00:33:34.615
- physics informed methods,
896
00:33:35.394 --> 00:33:37.575
- and and a lot of statistics and uncertainty
897
00:33:37.634 --> 00:33:40.934
- quantification. So all, you know, really important ingredients.
898
00:33:41.554 --> 00:33:42.035
- And,
899
00:33:42.355 --> 00:33:43.095
- then familiarity
900
00:33:43.394 --> 00:33:43.894
- with
901
00:33:44.630 --> 00:33:47.190
- pretrained models is increasingly relevant, and that could
902
00:33:47.190 --> 00:33:50.409
- come from language or image or video. So,
903
00:33:50.710 --> 00:33:52.409
- you know, many of these
904
00:33:52.789 --> 00:33:54.089
- same ideas and architectures,
905
00:33:55.589 --> 00:33:57.750
- kind of transfer into this domain and and
906
00:33:57.750 --> 00:33:59.289
- and and so those experiences
907
00:33:59.994 --> 00:34:00.974
- are all relevant.
908
00:34:03.034 --> 00:34:04.394
- I I guess what I've seen is,
909
00:34:05.994 --> 00:34:07.214
- it's not just about
910
00:34:07.674 --> 00:34:09.514
- one one skill that sort of makes the
911
00:34:09.514 --> 00:34:12.315
- difference for us. It is the collaboration across
912
00:34:12.315 --> 00:34:15.614
- physicists and simulation engineers and machine learning researchers,
913
00:34:16.050 --> 00:34:17.190
- infrastructure engineers,
914
00:34:17.570 --> 00:34:18.469
- you know. And and
915
00:34:18.930 --> 00:34:20.949
- from my own experience, often, you
916
00:34:21.410 --> 00:34:23.450
- know, there there's a different language there eve
917
00:34:23.570 --> 00:34:25.250
- even for the for the same problem that
918
00:34:25.250 --> 00:34:27.489
- you're trying to solve. And so bridging the
919
00:34:27.489 --> 00:34:29.930
- gaps between all these different experts is where
920
00:34:29.930 --> 00:34:31.590
- a lot of the value is created.
921
00:34:32.264 --> 00:34:35.385
- And so, you know, I think, of course,
922
00:34:35.385 --> 00:34:37.545
- being an expert in one area is is
923
00:34:37.545 --> 00:34:39.944
- absolutely the foundation of your credibility, and there
924
00:34:39.944 --> 00:34:41.464
- are, you know, the areas that I've talked
925
00:34:41.464 --> 00:34:43.864
- about there. But, also, I would advise you
926
00:34:43.864 --> 00:34:44.364
- know,
927
00:34:45.150 --> 00:34:46.609
- try and develop that familiarity
928
00:34:46.909 --> 00:34:49.329
- and understanding across the full scope
929
00:34:49.710 --> 00:34:52.269
- because the people that can kind of really
930
00:34:52.269 --> 00:34:54.190
- thrive are those that can have a meaningful
931
00:34:54.190 --> 00:34:54.690
- conversation
932
00:34:54.989 --> 00:34:57.170
- both with a a maybe a CFD engineering
933
00:34:57.835 --> 00:34:59.914
- engineer in the morning or or a system
934
00:34:59.914 --> 00:35:01.755
- architect in the afternoon and and and and
935
00:35:01.755 --> 00:35:03.055
- join those dots up.
936
00:35:04.315 --> 00:35:04.815
- And
937
00:35:05.114 --> 00:35:06.394
- one of one of the things I've loved
938
00:35:06.394 --> 00:35:08.635
- about being at PhysicsX is we we hot
939
00:35:08.635 --> 00:35:10.394
- desk here. And, you know, every day I
940
00:35:10.394 --> 00:35:11.775
- find myself sitting someone,
941
00:35:12.119 --> 00:35:14.359
- sitting next to someone with a really deep
942
00:35:14.359 --> 00:35:14.859
- expertise
943
00:35:15.239 --> 00:35:17.739
- in in something fascinating that I've learned from.
944
00:35:18.199 --> 00:35:20.839
- So I suppose my overall advice there is,
945
00:35:20.839 --> 00:35:23.339
- you know, be curious and be open minded.
946
00:35:24.515 --> 00:35:26.114
- Yeah. You never know where where the next
947
00:35:26.114 --> 00:35:28.135
- opportunity will come from or the next idea.
948
00:35:28.994 --> 00:35:30.675
- And and and there's a good chance the
949
00:35:30.675 --> 00:35:33.394
- next best idea for you will come from
950
00:35:33.394 --> 00:35:36.515
- someone else. So, yeah, stay receptive to people
951
00:35:36.515 --> 00:35:37.574
- around you and,
952
00:35:38.820 --> 00:35:40.599
- you know, always ask for feedback,
953
00:35:41.300 --> 00:35:43.780
- because it's it's really the best gift that
954
00:35:43.780 --> 00:35:44.599
- you can get.
955
00:35:45.059 --> 00:35:46.420
- So I think that would be that would
956
00:35:46.420 --> 00:35:47.320
- be my advice.
957
00:35:48.099 --> 00:35:49.619
- Well, it sounds I mean, it sounds like
958
00:35:49.619 --> 00:35:50.920
- a fascinating field
959
00:35:51.234 --> 00:35:52.054
- that you're in.
960
00:35:52.914 --> 00:35:54.695
- And, you know, especially for,
961
00:35:55.074 --> 00:35:56.914
- you know, for for for a physicist or
962
00:35:56.914 --> 00:35:57.494
- a mathematician
963
00:35:57.875 --> 00:35:59.875
- or, I suppose, even with you, a a
964
00:35:59.875 --> 00:36:00.375
- chemist,
965
00:36:01.474 --> 00:36:03.474
- that, you you know, there that there are
966
00:36:03.474 --> 00:36:03.974
- opportunities.
967
00:36:05.059 --> 00:36:06.739
- You know, I suppose AI gets a a
968
00:36:06.739 --> 00:36:09.139
- lot of bad press these days, but,
969
00:36:09.860 --> 00:36:11.960
- it sounds like it's really opening up,
970
00:36:13.460 --> 00:36:14.820
- a a a whole new world when it
971
00:36:14.820 --> 00:36:15.460
- comes to,
972
00:36:16.099 --> 00:36:16.840
- to design.
973
00:36:17.824 --> 00:36:20.704
- Yeah. You you absolutely. And and and, you
974
00:36:20.704 --> 00:36:21.344
- know, I think,
975
00:36:21.904 --> 00:36:24.385
- the the AI in the physical space is,
976
00:36:24.385 --> 00:36:26.385
- you know, is is quite a different beast,
977
00:36:26.385 --> 00:36:28.644
- and it's it's really accelerating.
978
00:36:29.184 --> 00:36:31.500
- And as you mentioned, you know, we we
979
00:36:31.500 --> 00:36:32.000
- we
980
00:36:32.380 --> 00:36:35.019
- have experts in in chemistry as well as
981
00:36:35.019 --> 00:36:35.519
- physics,
982
00:36:36.460 --> 00:36:36.960
- and,
983
00:36:37.739 --> 00:36:40.800
- you know, bringing bringing bringing those experts together
984
00:36:40.860 --> 00:36:42.380
- as well as, of course, the the kind
985
00:36:42.380 --> 00:36:43.280
- of fundamental
986
00:36:45.425 --> 00:36:47.585
- AI behind the under the hood, you know,
987
00:36:47.585 --> 00:36:49.285
- that's kind of where the magic happens.
988
00:36:50.545 --> 00:36:52.704
- Well, that's great. Thanks, David. Thanks so much
989
00:36:52.704 --> 00:36:54.864
- for, for joining us today and,
990
00:36:55.344 --> 00:36:57.585
- talking about your career and your work at,
991
00:36:57.905 --> 00:37:00.320
- PhysicsX. It's been a pleasure. Thanks very much,
992
00:37:00.780 --> 00:37:01.280
- Hamish.
993
00:37:09.019 --> 00:37:10.860
- I'm afraid that's all the time we have
994
00:37:10.860 --> 00:37:12.160
- for this week's podcast.
995
00:37:12.635 --> 00:37:15.914
- Thanks to Dave Wheater of Physics x for
996
00:37:15.914 --> 00:37:17.295
- an enlightening discussion
997
00:37:17.675 --> 00:37:19.775
- of how large physics models
998
00:37:20.074 --> 00:37:21.375
- are making cars
999
00:37:21.675 --> 00:37:22.494
- more efficient.
1000
00:37:23.355 --> 00:37:26.554
- I'm Hamish Johnston, and our producer is Fred
1001
00:37:26.554 --> 00:37:27.054
- Isles.
1002
00:37:27.699 --> 00:37:30.260
- The music for this episode is called one
1003
00:37:30.260 --> 00:37:33.320
- three seven, and it was composed and performed
1004
00:37:33.619 --> 00:37:34.599
- by the physicist
1005
00:37:35.059 --> 00:37:36.119
- Philip Moriarty.
1006
00:37:36.980 --> 00:37:38.840
- We'll be back again next week.