Real-Time Computer-Aided Optimization
Interview with Gregory Roberts – Flexcompute
Could you start by introducing Flexcompute’s work, outlining what you presented at CDFAM in NYC, and explaining how it differs from or expands on your recent presentation in Amsterdam?
Flexcompute’s goal is to build high performance, integrated physics simulators and tools that allow hardware innovation to be as fast and nimble as software. Through GPU accelerated solvers that can work together, multiphysics and engineering challenges can be both quickly and accurately modeled.
Combined with advanced design and layout tools, these models become actionable and Flexcompute enables rapid iteration and hardware development through software. In our CDFAM NYC presentation, we will talk more concretely about this process and look at examples of how we can power advanced and robust electromagnetic design via automatic differentiation and gradient-based optimization.
From a user’s standpoint, your GPU‑native CFD combined with adjoint and inverse design promises faster iteration and broader exploration. What specific gains should customers expect (e.g., wall‑clock time, cost per design, fidelity), and how does the method integrate with existing CAE workflows and data formats?
Customers should expect reduced iteration times when designing with Flexcompute’s tools. Not only does each simulation run quickly on our servers, but through larger simulations that can capture not just devices but also their environments accurately and robust optimization techniques, we can reduce the simulation to fabrication gap. Via automatic differentiation (auto-diff) and the adjoint method, we allow customers to optimize arbitrarily large and complex degrees of freedom with the additional cost of just 1 simulation per iteration. This makes gradient-based optimization an accessible, cost effective, and powerful design tool.
Our mature suite of optimization and analysis tools allows customers to shift more of their design cycle from hardware to software. Less hardware iteration is a huge win for customers in both time and cost. We support import and export through different data formats and also integrate layout information from foundries into our tools. For example, PhotonForge allows customers to directly set up simulations for a variety of PDK and photonics technology stacks.
Can you outline the data flow from setting design objectives through to the final optimized configuration, and where GPU acceleration has the greatest impact on reducing iteration time?
The short answer is a user simply needs to be able to set up their simulation(s) and write an objective function in python describing their figure of merit or loss function. This can be an arbitrarily complex function of different geometries and multiple simulations. Via auto-diff in Tidy3D, we can automatically compute the gradient of this objective function, allowing the user to employ a gradient-based optimizer of their choice to carry out the design.
As a specific example, let’s say you are designing a photonic component that splits power from an input waveguide into two output waveguides. The first step is building a simulation with the proper inputs (sources), outputs (monitors), structures (waveguides, substrates, background refractive index, etc.) and configuration (mesh specification, run time, etc.). Second, you can parameterize a structure that does the splitting, let’s say a polygon with controllable vertex points. Finally, you would write an objective function to capture your device performance. In this case, it may be the sum of the power into the two output waveguides. Since we have done the hard work to integrate with an automatic differentiation library, with one additional line, you can compute the gradient function for your objective function with respect to your design parameters (i.e. – the polygon vertices). With this gradient, you can iterate on your design, improving performance until convergence. This design process is iterative, meaning we end up running many simulations on the way to the final device. However, with GPU acceleration, this process becomes very fast and optimization loops that normally would take days can be done in a few hours even for large scale problems.
What technical challenges have you encountered in applying gradient-based optimization to large-scale aerospace or photonic design problems, and how have you addressed them?
Designing devices that are robust against fabrication inaccuracy and understanding how fabricated devices line up with their simulated counterparts are critical challenges. We are building tools and working with foundries to understand how to solve these problems via robust design, building fabrication models, and using gradients to understand sensitivities to a variety of process parameters and perturbations.
For a company unfamiliar with Flexcompute’s software, what factors or types of challenges should they look for when identifying an initial problem to address using your tools?
We aspire to help companies, individuals, and universities with a large variety of design and simulation problems. Due to the speed of the solvers, Flexcompute is uniquely suited for large scale modeling and design. However, with the integration of optimization, multi physics, and layout tools along with a simple-to-use GUI and python API, we believe any type of design problem can be tackled with our software suite.
What do you hope to share with and learn from the CDFAM audience about the adoption of GPU-native computer-aided optimization in industry workflows?
I hope to share with the CDFAM audience what Flexcompute can offer in the areas of physics simulation and design tools as well as a few goals we have for the near future. I’m curious to learn more from the audience what their pain points are with tools they are currently using and what critical problems they haven’t been able to solve yet. It would be great to understand more about challenging industry workflows and what types of designs people are interested in pursuing.
