0:03 Hey, good morning everybody. So I’m going to start our presentation by flex compute titled the real time computer aided the optimization and how GPU native simulation changes the industry. My name is Chi-Chi Wong. I’m a professor in aerospace engineering at the Massachusetts Institute of Technology and I co-founded Flex Compute a little bit more than 10 years ago. So today I’m going to present the vision of how flex compute aims to change how we innovate in hardware and through computer aided optimization and the mono is going to give us specific examples in electromagnetics of how we are doing so right now.

0:56 Why is hardware innovation slower than software? So when I was a kid, I have been reading various science fiction novels and according to these novels, by today 2025, all of us should be able to be taking vacations on the moon or Mars and a lot of us should have flying cars to go around not being stuck in traffic. Somehow this didn’t happen, right? Instead something else happened.

1:26 What are these something else? I didn’t have internet when I was young. Now we do. We didn’t have social media when I was young. Now we do. And the AI was not a thing when I was young and now we have AI. So these are software related innovations that are happening at a much faster rate than was envisioned when I was young. While hardware, aerospace particularly and others like energy, transportation, it’s much slower than what we have expected when I was young.

2:04 I have been thinking why that has been the case. So I was in Stanford University for my PhD in aerospace engineering and then I was in MIT for the last 16 years as a professor also in aerospace engineering. I have been observing like all the best students, PhD, PhD students in aerospace engineering, they did not stay in aerospace engineering when they graduate. They went to places like Google, Facebook, now open AI.

2:41 And why is that the case? It’s the case for a lot of reasons, but the fundamental reason is the following. So if they join open AI today, if they work very hard, have very good ideas, with a little bit of luck, their product, what they make may be able to influence millions of people in a year or two. Okay? But if they go to aerospace engineering at least the 10 years ago that was the case you’ll be very lucky if you what you work on impacts a lot of people in like 10 years.

3:18 So why is it the case? It’s because working hardware is risky, right? A lot of people lost their lives testing very innovative ideas. It cost a lot of money and it just takes a long time. So this is the fundamental reason why a lot of good people they don’t want to work on hardware they want to work in software and this is what we want to change as flex compute today so the dream of why we founded flex compute was we want to make hardware innovation innovating in like aerospace to be as easy as innovating in software for the past few decades we want to make it as easy in the next few decades so I’m going to go really quickly through what we have already achieved because Montre is going to give us more.

4:05 Basically we developed a GPU solvers that accelerates simulation by effect of 100. That was a 10 years ago. Back then nobody was doing GPU for simulations and while at the same time reducing computational cost and today our two products flow 360 and tidy 3D are being used by a wide variety of industry partners doing very innovative stuff. If you look at the logos, they are probably the most innovative partners in the aerospace and electromagnetic industries.

And something that we are specifically targeting for aerospace engineering is that the traditional way of designing aircraft and the spacecraft is really difficult. As long as you move past the conceptual design phase, the disciplines have very big barriers. What we are trying to do is using int what we call intelligent geometry to unify these disciplines to break the barriers between disciplines and really enable innovation to happen in aerospace industry.

5:15 And Munch is going to give us a different example of enabling people to innovate in electromagnetics. Thanks for the introduction, Chichi. Yes, so my name is Mchio. I’m leading the development of the DY 3D solver flex compute which is as Chichi explained focused on solving electromagnetic equations as well as some other coupled motif physics. And I realize that we’ve seen a lot of mechanical engineering for example we haven’t seen any electrical engineering here.

5:47 So I still hope that this will be useful for people in the broader ideas of what we’re doing. But as a bit of a background the some of the markets the main markets that we’re targeting or the main big market is photonic integrated circuits and some specific applications include optical interconnects. This is probably the biggest chunk which is becoming ever bigger actually because of the everinccreasing needs of for data centers specifically.

6:16 So AI has is now in many cases bandwidth limited actually and the optical interconnects are moving closer and closer to the to the electronic chips. There’s a lot of innovation happening there both from big companies and from startups that are by now already turning into big companies themselves. But other applications also include lidar which is also obviously very hot right now for self-driving cars. And also quantum computing we have some customers a lot of the basically doesn’t matter what quantum technology exactly you’re pursuing you’re always going to be using electromagnetic waves to either use as cubits or to access or connect your cubits.

7:04 So in this photonic integrated circuit domain there are design challenges on various different levels as is true for every industry. In general I’ve grouped them here into several different categories. Some are on the device level where individual components need to be simulated optimized and also made in a way that is robust with respect to the foundry processes. So chips generally have very very robust very specific requirements and every different foundry has different fabrication requirements that are very stringent.

7:41 And making the designs that can meet those requirements while still being very very optimal is actually very challenging. On the system level design once you have the individual components you’re still not done. Obviously, you have to be able to lay out you know, giant circuit of millions of these components. Make sure that everything works right. Make sure again that it’s also optimized and also there’s a lot of communication here that needs to happen.

8:11 Actually, that brings me more like to the workflow level. This whole process couples a lot of teams both or a lot of people within a team a lot of teams within a company and also there’s a lot of communication that needs to happen between let’s say the fabs and the the designers in a given company so there’s challenges across the board sorry I’m I’m going to focus on the first part here for the exact case study of what I want to I want to cover my mouse but I don’t know if that’s a good idea.

8:48 Oh, there we go. Yeah, I’m going to focus on the first part of the device level simulation optimization as a case study of of a recent innovation that we’ve introduced in this specific domain. And that’s something we call invoice design for everyone. Now I realize that in mechanical engineering people are quite familiar with topology optimization and invoice design. The reason we call it for everyone is because in the photonics industry actually that was not the case.

9:15 So topology optimization is not widely used. And part of the reason are those stringent requirements of the fabrication processes. So this whole device that I’m showing here is actually on the scale of a few microns. So that means that the features of the topology optimized structure are on the scale of tens or up to 100 of of nanometers. So this is very very hard to fabricate in a reproducible way over millions of devices.

9:39 And because of that inverse design topology optimization has generally infotonics been very interesting on the academic level for decades but pretty much not used at all in industry I guess for two reasons. One is ensuring this manufacturability and reproducibility. The other is also it’s it’s been a bit intimidating for people. It kind of feels like you know if you did your PhD in that domain you can do it but otherwise there really aren’t any tools that can do everything for you automatically if you if you’re like a design engineer but you you don’t have any specific skills related to input design.

10:15 Yeah so by inverse design generally we mean being able to search a relatively large parameter space. So a large number of parameters automatically let the computer figure out what’s best. But also it has to be constrained to the fabrication specifications. And typically or and also in what I’ll be presenting right now, typically this would use some sort of gradient-based optimization because if you have too many parameters, you need some way to navigate the parameter space and specifically the adjint variable method or automatic differentiation or it’s a related concept I guess is the the way that this can be done efficiently.

10:55 So gradients can be computed efficiently. Yeah. So in short the legacy invoice design has been quite complicated to implement yourself u hard to control on what is going to be produced and whether it will fit the specifications and slow using traditional computational methods. As GG pointed out we’ve definitely fixed the slow part of it because the initial offering of the company essentially well the the initial fundamentals of the company were developing GPU native solvers.

11:25 So compared to traditional tools you know you can get orders of magnitude speed up. And then we’re also solving the other two problems of legacy invoice design and that’s why we call it invoice design for everyone. So I can skip over this and go over a specific example. So for one component for example that is very common in these types of devices is a grating coupler in which you have light coming from above and you want to couple it on chip.

12:05 So this is basically crucial to couple in and out to optical fibers let’s say to couple across chips across racks and so on. So if we can compute the gradient information so this is actually a parameterized design so that’s even nicer. So here you you’re you’re not doing topology optimization per se in that your parameters are actually the primitives of just the spacings between the separate teeth of the device.

12:20 But yeah so if you compute gradients efficiently that’s great for two reasons. One is right away you get sensitivity analysis. You know how how much the let’s say the width of each tooth affects your objective function or the coupling efficiency and you know where to focus on. But then even better you can actually plug this into a optimization routine do some iterations and get a much better design than your starting design.

12:52 And then I think one of the innovations that I’m most excited about essentially is that we’ve we’ve made simulations largely different differentiable by which I mean that setting up a simulation is all you need to do to be able to also to compute gradients automatically. So this try to do this again with the mouse. Let’s see. So, so in Python for example, in our Python interface, what you need to do is define some functions that maybe create the teeth based on some primitive parameters like the widths and the gaps creates the simulation in tiny 3D.

13:29 Now all this is pseudo code obviously but the point is that this simulation creation most of the simulation components are can be tracked in our automatic differentiator such that the whole chain of parameterization simulation getting the results which is actually typically done through our through a code to our web.run function. So this actually uploads the simulation to a cloud. The results get you know it runs fast on our hardware.

13:57 The results get downloaded. This is also part of the computational graph. And then finally, you can have as complicated as you want objective function figure of merit. You can even combine data across multiple simulations. Everything is going to be tracked. So you can do like a you can have a Monte Carlo run of multiple simulation as part of your whole objective function. You can then as easily as calling this function to get the results you can call a gradient function that will also get you the gradient of the figure of merit with respect to the input parameters.

14:30 So that’s why we also call it sometimes the pytorch of photonics because if you’ve ever done a a machine learning model it probably you’ve used pytorch but even if you’ve not what you do there is you define the model you define the computational graph and you don’t have to worry about how how to compute gradients. Basically anything you define you can call grad on and you can get the gradient and you you can optimize it.

14:51 And it’s exactly how it works in 3D but with differentiable simulations and in machine learning essentially automatic differentiation is called back propagation. It’s the same thing reverse mode automatic differentiation specifically. It’s the very efficient way to compute gradients. It’s it’s probably not an overstatement to say that back propagation has enabled AI. And yeah so we provide simulative interface in in as I said in Python it’s basically equivalent to defining a simulation you can then compute gradients we also provide a GUI interface for that and then importantly we also provide essentially readym made functionality that you can define fabrication constraints like minimum radius minimum distance between features and so on that again you just only define these primitive parameter and everything will be tracked and included in the optimization.

15:41 And also we have several different algorithms. So you topology optimization is certainly one thing you can do but shape optimization. So for example moving these feet around without really going free form can can often offer often produce much better designs in terms of fabric and also level set optimization. And this is a quick example of how the teeth are moving around. And I’m going to skip to a topology optimization also example here again.

16:10 So remember this is on the micron scale. So but the objective is to optimize a meta service to have to focus the light injected from below to look like the flex compute logo which which is achieved here. And finally we also have another tool that we call tidifab which can actually and here we really work with the individual foundaries to get statistics. This this is usually proprietary information but when we have collaborations with foundaries we can incorporate these statistics into our tools for specific fabrication process on various uncertainties that one might have in the fabrication process that can also be again incorporated in the simulation and in the optimization such that maybe you won’t get the most optimal design possible but you’ll get something that when you fabricate you know you will get very close to the target results while you know if you include these things.

17:09 It might look great in simulation but then in practice it will not be very good. So going back to the challenges yeah I I I I had to focus on the on the first part the device level but yeah we have multifysics plus fast optimization of several different solvers that are GPU native and then the automated optimization on the system level design we have a design and layout tool that actually wraps on top of all of that and also facilitates the working with foundaries.

17:37 And finally on the workflow management tool we are also cloud native and so we also we have a lot of the u perks that come with that in terms of collaboration. That’s it. Thank you. To learn more about the CDFAM computational design symposium series, to see the archives of previous presentations, and to learn about future events, visit CDFAM.com.