YT YouTube2Text

YouTube2Text → Transcripts → AI for Design Optimization - Namwoo Kang | Deep Dive

AI for Design Optimization - Namwoo Kang | Deep Dive — Transcript

by Jousef Murad | Deep Dive · 3,646 words · 584 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Yes. Hello. I'm Professor Namukang from
0:02theQAIST smart design lab. The topic
0:05I'll be presenting today is AI for
0:07design optimization.
0:09I'll be talking about how we can use AI
0:12to make design optimization smoother.
0:15And for a case study, I'll be showing an
0:17example of wheel design which our lab
0:20has been consistently working on.
0:24To briefly introduce my background, I
0:26actually studied different fields for my
0:28undergraduate, masters and PhD degrees.
0:31My undergraduate degree was in
0:32mechanical engineering. So I have a
0:34similar background to many of you. After
0:36that, I studied technology management
0:38and then went to Hyundai Motor Company
0:39to do vehicle development research. Then
0:42for my PhD, I studied what's called
0:43design science, which is about
0:45interdicciplinary design. So it's a
0:47process where various fields like
0:49mechanical engineering, computer science
0:52and design come together to design.
0:55Currently I'm conducting research with
0:57about 20 masters and PhD students in the
1:00smart design lab. And from there we have
1:03a spin-off company called Nania Labs
1:06which is productizing and
1:07commercializing our work. It's made up
1:09of about 30 people.
1:12So these two groups have been
1:13consistently conducting research on
1:15generative AI based design.
1:19So first I'm going to talk a bit about
1:22why we should use AI for design
1:23optimization.
1:27First of all when we talk about design
1:28optimization we often refer to surrogate
1:31based design optimization. This is about
1:33performing design optimization based on
1:36some kind of simulation and the process
1:38is as follows. First you define the
1:40design optimization problem. Defining
1:42the problem means deciding what the
1:44design variables are, what the
1:46constraints are and what the objective
1:48is. Once those are determined and the
1:50design variables are set, you change
1:53those variables and perform CAE
1:55analysis. You then collect the key
1:57analysis data. Using that collected
1:59data, you typically use traditional
2:00machine learning methods to create a
2:02predictive model. You create a surrogate
2:05model and based on that model you use an
2:07optimizer to find the optimal value the
2:10value with the best performance.
2:13Now this type of design optimization has
2:15been done a lot but there are some
2:16fundamental problems some limitations.
2:19First when you optimize a shape you
2:22usually start by parameterizing it.
2:24However the shapes we're trying to
2:26design are often very complex and not
2:29easily parameterized. So one limitation
2:32is that it's very difficult to formally
2:34parameterize these complex shapes.
2:37Second, even if you create a surrogate
2:39model, if you have a very large number
2:40of input variables, if the input
2:43dimension is very high, it's difficult
2:45to create a predictive model with
2:47traditional machine learning models.
2:48That's another problem.
2:52And finally, in such a highdimensional
2:54design problem, finding the optimal
2:56design means it's also very difficult to
2:58find the global optimum. That's another
3:00limitation. So in a nutshell, real world
3:05industry problems are very
3:07highdimensional. So it's difficult to
3:09solve all problems with SPDO. That's the
3:11issue. So can AI address these
3:14limitations?
3:15Well, if we take a look, AI has many
3:18advantages, but its greatest strength is
3:20its ability to compress this original
3:23highdimensional space into a very
3:25lowdimensional latent space. And because
3:27this compression is possible, we can
3:29extract features and uh reduce the
3:32dimensionality. So what feature
3:34extraction ultimately means is that
3:36instead of manually parameterizing a
3:39complex uh object, the AI can
3:41automatically extract the important
3:43features and then we can perform
3:45surrogate modeling based on those
3:47features. And dimensionality reduction
3:49means that instead of optimizing in a
3:51highdimensional space, we can optimize
3:53in the reduced lowdimensional space,
3:55allowing us to find the value we want
3:57much more quickly. And when that
3:59happens,
4:02um, we can think of a framework for
4:04performing design optimization using
4:06this AI. You can think of it as being
4:08broadly composed of three modules. We
4:10have generative AI, predictive AI, and
4:12optimization AI. Ultimately X represents
4:15the design, Y represents the engineering
4:17performance and uh the ZV valueue
4:20represents the reduced latent space. So
4:23the generative model represents the
4:25mapping from Z to X. So X becomes the
4:27function G of Z. The predictive model
4:30represents the mapping from X to Y. So
4:32we're creating a function where Y is f
4:35of X. Ultimately by approximating these
4:38two functions, we can leverage them for
4:40optimization. There's an iterative
4:41optimization method and a generative
4:43optimization method. With iterative
4:46optimization, you essentially connect
4:47the generative and predictive models in
4:49sequence. You generate X from Z, predict
4:52Y using that X, find the optimal Z, and
4:55then input that Z back into the
4:56generative model to obtain the final X.
4:59As for the generative approach, you can
5:01simply think of it as conditional
5:03generation. You provide the generative
5:05model with your desired target
5:06performance as a condition and it
5:08generates the optimal X. So in the end
5:11the key takeaway is that the biggest
5:13advantage AI offers over traditional
5:15design optimization is that uh the
5:18optimization is performed in the reduced
5:20dimensional zpace. That's how we can
5:22summarize it. So then how can we apply
5:25this concept more uh to actual industry
5:29or real world scenarios? When we think
5:31about that, well, I'd like to try and
5:33redefine the problem once more. So, in
5:36the case of AI, since it's multimodal,
5:38the input can actually be any
5:40combination of 1D, 2D, and 3D. And the
5:42output can also be any combination of
5:441D, 2D, and 3D. So, while there are many
5:47possible scenarios, we can group them
5:49into broader categories. Are we trying
5:52to predict 3D, generate it, or optimize
5:54it? or predict 2D, generate 2D, optimize
5:572D or predict 1D or uh optimize 1D. We
6:01can group them into these larger
6:03categories for now. These are the eight
6:05key scenarios. I once wrote a short
6:07paper to try and explain this a bit more
6:09simply.
6:12So if we look at the 3D design scenarios
6:15from a prediction perspective, what we
6:17can do is you can think of this as
6:20referring to the inference stage after
6:22training. During inference, a 3D shape,
6:24a CAD model will be input along with
6:27conditions like load or boundary
6:29conditions. When those are input, the AI
6:32can predict a 3D field or a 1D scalar. A
6:34scalar could be maximum stress or
6:36maximum temperature. And a 3D field
6:38refers to all the nodal values. Also,
6:41depending on whether we're predicting
6:43statically at a specific time or
6:44dynamically and transiently at each time
6:47point, we must use the most suitable
6:49model. That's how prediction works. For
6:51generation, if we consider conditional
6:54generation, the conditions would be our
6:56desired design requirements. When those
6:58requirements are input, we need a model
7:00that generates a 3D shape. Then with the
7:03generative and predictive models, as I
7:06explained earlier, whether these models
7:08operate iteratively or are trained
7:10simultaneously for conditional
7:12generation, they can produce an optimal
7:14design. You can think of this as the big
7:16picture.
7:18In fact 2D is the same. Usually in the
7:21field when problems are difficult to
7:22solve in 3D we can u project the 3D into
7:262D. This could be a depth map a
7:29cross-sectional design or a 2D image
7:31from a specific view. Then similarly the
7:34predictive AI will predict field values
7:36scalar values static values and dynamic
7:39values for generation. It's often
7:42difficult to generate directly in 3D due
7:44to a lack of data. So instead of
7:46generating 3D directly, we generate at
7:49the 2D level which could be a depth map
7:51or multiv- view images and then
7:53reconstruct them back into 3D. Then
7:55again the predictive and generative
7:57models can be combined to create an
7:59optimal 2D design and then an optimal 3D
8:01design.
8:04Then 1D is yes even simpler. We can
8:08create a 1D prediction model and a 1D
8:10optimization model and then combine
8:12them. For 1D generation is simpler than
8:153D or 2D. So I've excluded it here.
8:19And um for 1D prediction, a technique we
8:21often use is multifidelity prediction.
8:24In the field, there isn't much
8:26highfidelity data, that is expensive and
8:29accurate data. So we often use
8:31lowfidelity data, which is cheap and
8:33abundant together. So we do a lot of
8:35research on how to smartly use both
8:37lowfidelity and highfidelity data for
8:39prediction.
8:42So as I mentioned earlier, these are the
8:44eight important key scenarios. The
8:46choice is made depending on the data the
8:48engineer has. In detail, as I mentioned
8:50briefly before, in prediction, there's
8:53scalar prediction, field prediction, and
8:56time series prediction. In generation,
8:59there's random generation, conditional
9:01generation, and assembly generation. For
9:04optimization, you can choose to do it
9:06iteratively or generatively.
9:08Furthermore, if you want to improve
9:10prediction performance, there are
9:11methods like multiffidelity prediction,
9:13methods that predict not just the result
9:16but also the uncertainty to show its
9:18reliability. And since there are limits
9:21to using just data, we can input the
9:24physics information we know and use pin
9:26physics informed neural networks. So
9:29various combinations are possible. So
9:32not all problems can be solved with one
9:34method. So we need to use these various
9:36combinations to suit each problem.
9:40So when this became possible, we created
9:42software to demonstrate this. This way
9:45we could generate thousands of 3D shapes
9:48using deep learning. Since deep learning
9:50can judge the similarity of shapes in a
9:52lowdimensional space, we could group
9:54them, color them, and predict their
9:56performance when you click them. We
9:58could cluster them together with similar
10:00performance or shape. And since the
10:03performance is predicted, we could
10:05filter out the ones that match the
10:07desired performance and we could compare
10:09the design we created with our previous
10:12design or we could input the target
10:14performance and immediately
10:16conditionally optimize the design.
10:21You can download the design plans and
10:23create detailed designs. You can also
10:25share your favorite designs with other
10:27designers for comparison and use. You
10:30can also combine existing designs to
10:32create new ones. This approach seems
10:35possible from a broader perspective.
10:38Let's take a look at how we've applied
10:39these concepts to wheel research at our
10:42smart design lab SDL. So, let's take a
10:45look at how we've applied these concepts
10:47to wheel research. It first started in
10:502018.
10:52Looking back from 2018 to 2025, there
10:55have been a lot of changes in the AI
10:57field. At that time, we could not have
10:59predicted that AI models would develop
11:02this quickly. And along with such AI
11:05development, we constantly thought about
11:07how we could further reduce engineering
11:10design. And as it turned out, about 11
11:13of the papers we wrote had research on
11:16wheels like this. So we combined those
11:18and we'll briefly introduce them. So the
11:21first research question was when
11:23designing 2D it is not just important
11:26from an engineering perspective but also
11:29aesthetically since it is a visible
11:31design. How can we combine engineering
11:33performance and aesthetics and make it
11:36this is the problem we solved. In fact
11:38in a way this was the very starting
11:40point of our research. I wrote this
11:43paper after presenting at the IDTC
11:46Institute of Mechanical Engineers in
11:482018. At the time, we were researching
11:51how to create a design that resembles a
11:53reference design while also being
11:55superior in engineering quality rather
11:58than just considering engineering
11:59performance.
12:01Based on that data, we developed a
12:03generative model which at the time was
12:05GAN to generate more references and then
12:08based on those references, we could
12:10iteratively create designs that were
12:12similar to the reference but also more
12:14engineeringly superior.
12:16So this was the beginning of our
12:18research and fitting for a foundational
12:20study. I recently saw it has been cited
12:23over 600 times and has become a
12:25well-known study in the field of
12:27generative design.
12:29Starting from this we kept thinking when
12:31generating we also have to consider
12:33diversity. So it's not about the
12:35quantity but how evenly distributed they
12:38are that's important. So our next study
12:40used reinforcement learning
12:44to learn how to explore and design the
12:46most diverse set of designs.
12:50Next, I mentioned aesthetics earlier and
12:54aesthetics are about emotion, right? Of
12:57course, if the reference design is good,
12:59you can create more aesthetically
13:01pleasing designs, but we can also
13:05evaluate how people feel about AI
13:07generated designs through online
13:09surveys. So, we had North American
13:11customers rank the generated designs and
13:14based on that we carefully developed a
13:17model to predict what kind of designs
13:19people like. When building this model,
13:21we didn't just use deep learning. We
13:23conducted in-depth research on which
13:26features to extract to better understand
13:28and predict wheel patterns.
13:32Then when an actual designer evaluated
13:34them, we found that with the designer's
13:36evaluation score on the xaxis and the
13:39predicted engineering performance on the
13:41y-axis, we could extract a parto front
13:44showing the trade-off. So as you go to
13:46the left, the designer likes it, but the
13:48performance drops. And as you go to the
13:50right, the designer dislikes it, but the
13:52performance improves. We expanded this
13:54into research that allows us to see this
13:56trade-off in a datadriven way.
14:00Now the second question is this time we
14:03always try to build a deep learning
14:05based surrogate model, but when you go
14:08into the field, the problem is that to
14:10build a surrogate model, you need X data
14:12and Y data. The Y data can actually be
14:15accumulated somehow even by manual
14:17analysis as long as you have enough X
14:19data. The problem is when you don't have
14:21X data which is very tricky. Having a
14:24person create diverse CAD models one by
14:27one is nearly impossible in terms of
14:28time and cost. So this is where we often
14:31get stuck. So how can we build a good
14:33surrogate model without X data?
14:36Ultimately we're in a situation where we
14:38really have to create synthetic data. So
14:40what we initially started with was
14:422.5DAD automation which is a process
14:45that uses computer aided design tools.
14:48Since we can already generate various
14:50designs in the spoke view as I showed
14:52earlier, we collected and trained on the
14:54cross-section image data of wheels that
14:56a company has to generate new
14:58cross-sections. Then by carefully
15:01utilizing CAD automation to combine the
15:03spoke views and cross-sections we
15:05created, we were able to generate
15:07various step files.
15:10And through CA automation, we could
15:12create a prediction model. And through
15:14that model and explainable AI, we could
15:17show which areas to focus on in terms of
15:19performance. So we were able to create
15:213D synthetic data like the following.
15:25Then once we have this diverse data, we
15:28can run various analyses, not just
15:30simple stiffness analysis like before,
15:32but by accumulating strength analysis
15:34data through impact tests. And based on
15:37that data, this is a case where through
15:393D deep learning, when a CAD or step
15:42file comes in, we can immediately use
15:44voxalbased or point cloud-based methods
15:46and through various representation
15:48methods predict desired information,
15:51field values or even maximum stress
15:53values.
15:54So ultimately we can know the location
15:56of maximum stress which originally
15:58required running a full analysis in real
16:00time through deep learning prediction.
16:05And going a step further then what
16:08if 3D prediction is possible we start to
16:10look a bit deeper. Uh when we tried it
16:13it predicts well with the data I have it
16:15predicts well but
16:17when I actually develop a new product
16:19and put it into the existing prediction
16:21model it doesn't predict well because
16:24it's impossible to gather and pre-train
16:26on the distribution of all possible
16:28situations. So if you look the data I
16:31have will come from multiple sources and
16:33each will have very different shape
16:34features. So we asked how can we train
16:38on these things to predict well even
16:40when a new target arrives.
16:43That's why we did domain adaptation. So
16:46by using only the common features of
16:48these wheels we can predict what we
16:50want. We conduct such research and also
16:53uh depending on the representation
16:54method various 3D analysis predictions
16:57are possible. For example, if I say I'll
16:59predict based on a mesh, a graph, then
17:02the prediction performance also depends
17:03on how that graph is created. So we also
17:06research algorithms that can pre-process
17:08the data to optimize the deep learning
17:10prediction results. So in a way to
17:12advance a single prediction model, a lot
17:15of consideration was needed from various
17:17aspects.
17:19And the third point, so in the previous
17:21step like this,
17:26if we designed based on 2D or 2.5D, we
17:29started to wonder if there was a way to
17:31directly optimize 3D designs without
17:34parameterization
17:36because in most real world problems,
17:38parameterization is difficult. So the
17:41method we've used a lot and are
17:43currently using is implicit neural
17:46representation. So it ultimately
17:48represents the three-dimensional or 3D
17:50shape by having a deep learning model
17:53predict the distance from any arbitrary
17:55point in space to the surface of the CAD
17:57model. Then in the case of 3D the
18:00network for representing it has a part
18:02where the coordinates go in the XYZ
18:05coordinates and if there's a latent
18:07space the part it predicts is the
18:09distance from this coordinate to the
18:11surface. So if I input any arbitrary
18:13point I want, it can predict the
18:16distance from that point to the surface.
18:17So we can reconstruct it based on that
18:19distance. So by changing the ZV value
18:22like this, we can reconstruct new
18:24designs.
18:27Then based on this we try non-parametric
18:31optimal design. So the design variable
18:34isn't a parameter but that latent space
18:37becomes our design variable allowing us
18:40to find a parto front where stiffness is
18:43maximized and minimized or weight is
18:45maximized and minimized. Um but it even
18:49works when trained with just 30 data
18:51points but in an extreme case with just
18:54two data points if you train the neural
18:56implicit representation well it
18:58interpolates well. So in a way even when
19:02data is scarce in the design field we
19:04could perform non-parametric optimal
19:06design like this.
19:10Next moving on to the fourth problem.
19:12What we want in the industry is to
19:14carefully and attentively consider very
19:16very detailed shapes.
19:19How can we optimize a design that has
19:21such realistic details that anyone can
19:23see?
19:25So to do it directly at the 3D level uh
19:29foundation models in 3D are still
19:32developing. So at some point they will
19:34certainly be usable but for now the 2D
19:37foundation model is so powerful that we
19:39are adopting a method to utilize that
19:42model as much as possible.
19:44So whereas in the past when we generated
19:47based on our previous 2D based research
19:49we only created black and white topology
19:51shapes like this now we use stable
19:53diffusion technology to first create a
19:56very photorealistic image.
19:59This image
20:03ah from that photorealistic image we can
20:06now use the shading to confirm the sense
20:08of depth the sense of depth.
20:12Uh this is let me first show you a video
20:14of how it's made. So when we created the
20:16software, we first generate from basic
20:19black and white patterns like this.
20:21Select the number of spokes and if you
20:24choose the desired shape, you then apply
20:26some styling to it. So if you input it
20:28as a text prompt, you can generate it.
20:36As soon as it's generated, it predicts
20:37the engineering performance like weight
20:39or stiffness and you can try it on an
20:41existing vehicle. This way, a designer
20:44even without an engineer. In fact, you
20:46could say the designer, engineer,
20:48salesperson and marketer can all gather
20:50in one place and plan a good concept
20:52vehicle for the future.
20:55So based on AI a somewhat photorealistic
20:58then as I mentioned earlier we first
20:59create a uh photorealistic image then to
21:03extract the uh spatial depth from it and
21:06create a 3D model we utilized a depth
21:08map after estimating the depth map we
21:10use that estimated depth map to
21:15reconstruct it into 3D that's the
21:17technology we developed because of this
21:20in the past when we made designs in 2.5D
21:23we could only create very flat designs,
21:25but now we can create 3D models that
21:27have that sense of spatial depth.
21:30So when a 2D image like this comes in,
21:33we can reconstruct it into a 3D mesh
21:35making full use of the 2D foundation
21:37model.
21:39And so we actually recently released a
21:41data set. So if anyone is interested,
21:44you can go to our website, our smart
21:46design lab data set website and download
21:48it. So it includes 2D RGB images, depth
21:52images, and these 3DS STL files, step
21:56files, and even the engineering
21:58performance data.
22:01And so since we can go from 2D to 3D,
22:05meaning that mapping is possible from
22:07the very start of 2D generation, we can
22:10create designs while considering their
22:113D performance. So while we generate 2D
22:14images, we do so while considering their
22:163D performance. And ultimately when
22:18these are created, it's not simply about
22:20making something that looks plausible to
22:22the eye. This is a case where we
22:24optimize while considering the
22:25engineering performance it will have
22:27when reconstructed into 3D.
22:32Yes. So now for the summary. So uh to
22:35recap uh at the very beginning I
22:37explained the reasons why we should use
22:39AI for design optimization and the
22:41limitations. To overcome those
22:43limitations, we uh classified the kinds
22:46of uh scenarios that can be implemented
22:50from a problem definition standpoint.
22:52And in fact, these eight scenarios are
22:54what we ultimately applied to the wheels
22:56in various ways. So we've applied them
22:58quite evenly to uh 2D problems, 2.5D
23:02problems, 3D problems, and so on. And it
23:04will likely develop very quickly in the
23:06future seeing the recent uh progress in
23:09models. So, it's expected to continue to
23:11become more advanced.
23:14Uh, that's all the material I've
23:16prepared.
23:17So, if you have any questions, I'd be
23:19grateful if you'd ask.

About this transcript

This page contains the full transcript of AI for Design Optimization - Namwoo Kang | Deep Dive by Jousef Murad | Deep Dive, generated from the public captions YouTube serves with the video. The transcript has 3,646 words across 584 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.