This episode of Designing the Future is brought to you by Altair.
Learn more about design generation, design exploration and design optimization, powered by AI.
The transcript below has been edited for clarity:
Jim Anderton
Engineering is applied science. The practical application of fundamental laws in disciplines like physics and chemistry rooted in the mother of all sciences mathematics. Now, today, engineers think not so much about numbers, but data. And today, data encompasses much more in engineering than renderings of parts and assemblies. High performance computing and artificial intelligence are taking the design process beyond the simple simulation of what if scenarios and are creating new and unique multi-directional data streams.
Data which must be analyzed, understood and acted upon to optimize engineering designs with speed at low cost. The possibilities are limitless, and they extend to every aspect of engineering in all disciplines. Joining me to explore this important subject is Fatma Fatma Koçer, vice president for engineering data science with Altair, where she and her team works on engineering, data science, strategy development, execution, investigating and applying latest technologies in the field, and providing feedback to alter software and supporting customer projects.
She's received a B.S. degree in civil engineering from the Middle East Technical University, Ankara, Turkey, and both MSC and Ph.D. degrees from the University of Iowa in Structural Optimization. Dr. Koçer has been recognized in Crain's Detroit Business 2019. Notable Women in STEM Report. Fatma, welcome to the program.
Fatma Koçer
Thank you so much, Jim.
Jim Anderton
Can we start with a sort of a generic concept to get everyone up to speed on optimization? Now, optimization used to be a very simple, iterative process when I was doing this, and that was we designed something, built a prototype, break it and go back and redesign it.
And we do that again and again in iterative way down there. How what does optimization mean today?
Fatma Koçer
It’s been a while that somebody asked me that question, so thank you for that. Optimization is actually the use of mathematical procedures to automate the objective of finding the design parameter values that meets a set of design requirements while minimizing or maximize using an objective function. I am familiar with the way you describe design optimization, but the goal in all in I should say true design optimization is to automate that process to find the values that minimizes an objective function such as cost, while meeting all the design requirements that the application needs to meet.
Jim Anderton
Now, in many design processes, you know, the matrix grows, so you almost grow exponentially as you work your way through the design process as problems crop up, which must be solved, or sometimes scope creep where the parameters are expanded by the demands of the customer or other downstream parties. Is, is how much of design optimization, I should say, is based on a careful definition of what the requirements are. I mean, at the beginning of the process, how important is that.
Fatma Koçer
The most important step? And one of the advantages I think, that design optimization brings to the table is the step to engineer has to think through the entire product lifecycle and formulate those design requirements upfront so that they don't creep in later on in the process when the design has matured. So as a design optimization engineer, for example, on, I was tasked with analyzing the lifecycle of a product and making sure that all the design requirements, all the operation requirements are included in the design optimization process so that we don't run into an issue later on. So it's one of the most important steps of the design optimization process, formulating the problem correctly and as a whole.
Jim Anderton
You mentioned overall lifecycle and today the overall lifecycle of a product now includes what happens to it at the end of its life. We're all very sort of environmentally sensitive now, and I know there's a large German manufacturer of luxury cars now that actually takes the vehicles back at the end of their life and disassembled them. And this has had interesting engineering implications, like the minimization of the number of different grades of plastics in use to simplify the process of the back end.
Now, if that ripples through back to the front end of the process to design, suddenly now the engineer doesn't have the luxury of being able to select an engineering resin to get the modulus they need. They may be forced to use a commodity resin because there are only a handful of available resins to use because of the recyclability sort of factor.
Is this the kind of thing that complicates design optimization these days, this end-of-life issue?
Fatma Koçer
It's one of the issues. And as an alternative, we actually do more of MOCA objective optimization in design optimization, meaning that you look at multiple objectives. In the past, we were more inclined to look at, you know, improving the strength of the design while minimizing the cost all the way. But now, more and more, we're using multiple objectives in search of the optimal design, including the ones that you mentioned in terms of recycling of the product or the comfort of the, you know, the products, the cost, the manufacturing of the product is also one of the biggest considerations nowadays.
Jim Anderton
Yeah, we have a new world now where computer aided engineering and computer aided design has allowed us to do things like highly optimized designs that were greatly simplified previously. So, you know, a triangulation of forces, you know, a Warren truss, you can always reach into the handbook solution and find a relatively simple structure that would handle those loads.
Now we have things like new materials with variable moduli and we have of multiple sort of options that are complicated. Is it still possible for engineers to start the optimization process with a handbook solution? Do they start simple and iterate toward complexity, or do they have to start in the middle someplace?
Fatma Koçer
That's a good question and I agree which computer aided engineering? We are being able to tackle much complex problems, both in terms of geometry and in terms of physics, just like we used to do single object optimization. Now we're doing more multivariate optimization. We used to do single physics because that's all we could handle, both in terms of our solution processes and in terms of our computational resources. But now we're doing more and more multi physics problem like field structure interactions, solid mechanics and electromagnets. So yes, the problems are getting complex, but we are engineers are our job is to find simple solutions to complex problems.
Jim Anderton
How much of this is cost driven? I know the automotive industry that I come from. Cost is really a primary consideration and there's many situations where it would be it would be useful to be able to use an exotic material, for example, a high strength alloy, but it's simply cost prohibitive to do so. So more weight must be engineered into a structure to make up for the strength that could be had easily with with, say, Iron Modulus material, for example, or a more sophisticated design.
Is it is is cost still king in most of the engineering disciplines that you study.
Fatma Koçer
Cost is the king and they're all cost you one are maybe in aerospace more risk driven and then the task to the engineer is within the limitations of the cost still find acceptable or feasible designs a design that meets all the requirements. And I think that's what we excelling as engineers what we are excelling a trade with cost human cost limitations.
How do you find type performing designs? Of course, there are some industries that cost is not as much of a driving factor, I would say, such as electronics and some consumer products. But yeah, it's of course an important piece of the puzzle.
Jim Anderton
You mentioned the aerospace industry. The aerospace industry has been revolutionized by things like composite materials and sort of a paradigm shift from metallic structures where we knew how strong it had to be, we knew what the safety factors were. So, we simply engineered to a given safety factor, perhaps tested a representative example just as a verification step. And then we moved on in things like composites, we have structures would operate by statistical sort of laws where it's, it's we bend it X number of million times and there is a resultant Y probability of failure and that we operate in a different sort of way of thinking about cycles to failure or cycles to that are reasonable within a sort of a design life expectancy. It's this is not something you can test by making something and breaking it out. Where does design, optimization and simulation fit into that new world where things operate with unique materials and unique ways?
Fatma Koçer
So actually, when we talk about new materials, such as composites, such as additive manufacturing, it's a real example, a field where simulation and data science converges because for some of those new materials to do not have enough understanding of the physics to use simulation fully, but we don't have enough data either to be able to drive properties solid or fit.
Well, we can use both of them combined, right? Our limited understanding of the physics with the limited amount of data that we have speaking, combined, combine and then, you know, still be able to design all safe, reliable products. So I think that's a good example of the convergence of simulation and data. For some of those, you'll be surprised that we actually have much more data then then we will.
We think there is because there is very high-end testing facilities in the nation, labs around the country. That is supercross is just testing these materials, whether it's composites or additive layered manufacturing, and provide the data to OEM, civil, aerospace or automotive. And there are we're doing data driven decisions in picking the right rate design parameters like design dimensions.
And that's where data science is complementing the simulation, the computer engineering world, as we know more and more about the properties of these materials, of course we'll be developing the physics that solves them and we may rely less on data. But this is a good point where more data and simulation helps with data. Of course, we can also understand the uncertainty and the confidence, confidence in the models, and it becomes more important with these new materials.
New physics combinations are in for especially new complex applications
Jim Anderton
So much of engineering is empirical and basically, it's many of the functions were derived from empirical observation. And it sounds like you're talking about a world which operates the other way around where the mathematics informs the design rather than the experience of the designer informing the mathematics. Is that is that a shift in the way and design engineers should think.
Fatma Koçer
With design and optimization? Yes, the mathematics drive the design decision making process, but it's to engineer that form leads to problem and it's to engineer that's going to be interpreting the results, making sure that they're meaningful, they're there, they're feasible. So it's a combination of engineers, expert knowledge and the data science. So one of the things, for example, that we're working in my team, we call it expert, immigration expert.
That's one of the AML solutions we're integrating into hyper works. And on the objective in expertise, AI is to emulate, to mimic experts decision making process. In automated design optimization, there are a number of design requirements that are not quantitative like cost but good quality to like the look and feel of the product. Whether the the behavior is, you know, favorable or unfavorable.
And so that means during the design optimization process, we would have to have the human in the loop, which breaks to automation, which makes the process much longer. So, using machine learning, our particular classification models, we take to the data from the engineer in how they had sort of judged the previous designs. We train an animal model that can be used in our optimization products so that the design can be optimized in one iteration.
So again here, once the engineer formulates the problem, it's the math that takes over, but really it's the engineers problem formulation that determines what the outcome would be.
Jim Anderton
Now, that's interesting. You mentioned that the of course, the traditional way to iterate to design success was often constrained by cost and by time. I recall many circumstances when perhaps we had four months to to formulate a finished design that was locked in for production and that distilled down to perhaps 18 or 20 testing cycles and whatever the product was.
At the end of that process, that was the design that was locked down for production. And many times you watch things go to production and think, Boy, we could have made that better. We could have pulled more cost out of it, we could have made it stronger, we could have made it more durable. But we ran out of time.
So in some cases, the more experienced engineer, the quote unquote better engineer in the design office was the one who could start at the middle or who could optimize the design on paper or as a rendering before the testing cycles began. So there are fewer cycles required to to optimize the design. You're talking about a mathematically driven world.
Is that over now? Are we a world in which basically the the the machine says that that the more experienced engineer has no specific advantage over the the mathematically driven, data driven newbie?
Fatma Koçer
No, actually, it's almost the opposite. All the expert engineers knowledge of the products is very important in the in engineering in terms of problem formation, what better designs can be find form better problem formulations. So if you're an expert, an engineer, your problem formulation mimics your design requirements much better than if you were a new engineer. But so this is another point where simulation converges with machine only because we can learn from the historical data the expert engineer has created and trained machine learning models and deploy these models to meet the younger generation of engineers so that they don't always have to start from the beginning.
It's almost like they would be inheriting some of the knowledge that the expert engineer has in the form of these machine learning models. One of our other projects of integrating machine learning and AI to our simulation software is this exact notion of using historical simulation data. So being able to recycle the past experiences and expertize that the company has.
And the goal for that is again what you have mentioned, which is designing better products faster. So if you can train animal models using your historical simulations, then animal models run really quickly so you can explore many more designs in much shorter amount of time. And this larger exploration will give you more alternative designs. And then once your commercial feel that's promising, then you can do the old school physics-based simulation, which usually requires more computational and time resources, and also do maybe one or two physical testing.
But that reduces to the expense of physical tests and end also, you know, computationally intensive physics based simulations. So this is another example of how Altair is converging on data science, machine learning, AI, rit, which see with simulations, stewards.
Jim Anderton
Fia CFD are examples historically were to a certain extent constrained by available computing power, and we've all seen circumstances where engineers design things, perhaps not optimally, but designed things that can be calculated. So we're looking at a world now in which we have high performance computing that operates. I mean, we're talking about, you know, exascale computing, you know, petaflops incredible, incredible speeds at this point is that has those shackles been removed? Is it at a point now where you can design the complexity and go ahead and use your physics?
Fatma Koçer
I mean, to say it's much better than when I started in the industry. We were at a point that if you run five simulations, the six simulation would be questioned because we were you could be using the resources. Right. But it's not anymore like that. I mean, people submit hundreds of simulations and be able to get the results within a really, you know, within acceptable time ranges like a day or maybe even a week.
So HPC is a very essential part of this process. All the barriers are much, you know, has been removed from the past, both in terms of the ability to access to these resources, especially with cloud services where the resources become expendable. For example, in our offering in all Tier one, both in terms of data storage in terms of job submission, in terms of resources like CPUs and GPUs, it's expensive.
All expendable, right? It's per paper usage. The more you need it, the more you have access to it. So you're not limited to your fixed boundaries. As you know, this is before. So yeah, people are estimating much more of these solutions are much more accessible, much more affordable. And of course, another aspect of it is with all the data that you're generating and then we're using for machine learning models, the simulation data management, that's also becoming a very, you know, even more important now than before, because there are all these going back and tapping the data that we've generated a year ago or two years ago, because that's what we can use to train animal models, deploy them for quick design explorations, quick design, you know, decision making processes.
Jim Anderton
And for the experience and knowledge that we're talking about, the very valuable experience used to be contained in the cranium of a good engineer. Then it expanded perhaps to the cumulative efforts of a team of engineers in a design office. Now we have a world which is cloud connected, and we're talking about high performance computing, which naturally suggests cloud connectivity.
What is the difference now in a world in which engineers can be connected instantly and in real time to other groups or other machines or other technologies around the world? Is this going to change, you think the way that we simulate to success?
Fatma Koçer
I think so, because cloud means collaboration. More collaboration, cloud means easier access to the information and to the data, to the physics based simulations. Cloud means expandability in terms of your usage. And to me after three, collaboration is a very important aspect because then we can merge, you know, the expertise that's in different parts of the world, that's in different parts of the teams and make design decisions, you know, to the considering the product as a whole.
So we're not limited to our silos anymore. We're not just working, for example, to meet the New Age objectives. We are cooperative with the team that's working on disability or the team that's working on fuel efficiency and doing a true multi objective design optimization.
Jim Anderton
It's it's interesting, we there was a time when some production technologies were invented and the existence of new things like C and C machine tools for example, then gave engineers ideas for designing products that could be made with this new technology. So does this have the way around? Sometimes it's the design requires a new technology to be developed simply to make it an exemplar.
So, for example, in the sophisticated forms of welding, for example, or adhesive bonding, we have something that is not new, but we think of it as new, and that's additive manufacture ring. So we have the ability now to make shapes that are not constrained by the traditional triangle of forces. And I know engineers, engineers are trained to think in rectilinear terms, in terms of Cartesian coordinates, and that sort of organic shapes are nothing more than, you know, line segments and planes, you know in a mesh shrunk down get it to an infinite level is are technologies like additive manufacturing is this going to change the way we design and the way we simulate to success, the fact we can make any shape anywhere?
Fatma Koçer
Yeah, it is it actually is changing and talking about, you know, engineers liking rectangular things. So before topology optimization, before topography optimization, all the stiffening members that you see in things like brackets or old drips that you see in stent parts, there are all these, you know, perpendicular members, you know, nicely spaced, equally space, I should say. But then topology optimization, type of optimization came along.
And now we start on, we start getting used to these organic structures that makes better use of the material, better use of the design space. So just like that with Additively manufacturing, we are now, for example, seeing these letters structures or structures that have, you know, repetitive pattern that is, you know, that wouldn't be able to be manufactured in any other ways.
Yeah, it looks unique, it looks different, but it's also expanding our design space, which means that we are being able to design even lighter, even better performing designs.
Jim Anderton
Recursion is something in nature that you see frequently, things like on a macro scale, and as you go down to a micro scale, you see repeating patterns over and over again. Is that something that that lends itself to this mathematical physics-based approach in design for that lightweight structures? You're talking about lightning holes within lightning holes within lightning holes.
Fatma Koçer
That's a good question. The pattern repetition is, you know, sometimes because these technologies are new and there are some limitations that they convert so that we would have to repeat the same pattern. But just the fact that these patterns are very light. You know, we're literally cutting holes in very much smaller scale than we did before.
It allows us to achieve lighter designs. But we were seeing that, for example, again, machine learning combined with this pattern increase creation, we don't have to be limited to the repetitive pattern. So we can change the pattern and even reduce more read, create more lightweight structures. Still meeting all the design requirements for this.
Jim Anderton
There's so much to talk about. We could go on we could do this for hours, I'm sure. But, you know, the clock must intrude. A final question for you. And this this is sort of it's near and dear to my heart, to many engineers. It's in the design process. It's because of the limitations traditionally and the number of times you can iterate toward success, you have to start with a good design and then use you simulate or test to success to a very good design or a great one if you're lucky at the ability to simulate quickly and in, in vast number of runs over and over, thousands, millions, perhaps at this point, does this mean now that it's possible to try things that you would not dare try from a design standpoint before? Can you throw something against the wall and see if it sticks? Can you try something crazy and just press the button there and stand back and see what miracle evolves?
Fatma Koçer
That's exactly where the convergence of simulation and machine learning is for us. Again, we can use historical simulation data to train models, and these animal models gives you the performance of a new design instantaneously, right? So that means we can try many more designs than before we can go crazy. Of course, we would have some indication of whether that that that prediction is reliable or not.
But as you have more data, you would be training with more variety and your your email model would be able to have confidence even in, you know, in crazier designs. Right. So this is this is the power of simulation and and machine learning, and we can be trained much more quickly. One of the all the limiting factors, all relying entirely based on physics-based simulations, was the amount of design exploration. But with the convergence that barriers also being lifted.
Jim Anderton
Fatma co-chair Altair, thanks for joining me on the show. And thank you for watching. See you next time on Designing the Future.
Fatma Koçer
Thank you so much. Thanks for having me.
Jim Anderton
And thank you for watching. See you next time on Designing the Future.