How GPU-Native Simulation Changes Engineering, from Consumer Electronics to Aerospace
Interview with Momchil Minkov + Qiqi Wang of Flexcompute
From aerospace to quantum photonics, engineers are facing increasingly complex design challenges that demand speed, accuracy, and scalability.
At CDFAM Amsterdam 2025, Momchil Minkov and Qiqi Wang of Flexcompute will present how Computer-Aided Optimization (CAO)—driven by inverse design, adjoint methods, and GPU-native simulation—is reshaping what’s possible in high-performance engineering.
Their talk spans everything from aerodynamics to meta-lenses, demonstrating how rapid iteration and goal-driven design exploration can unlock new frontiers in both hardware and software.
In this interview, they explain how CAO differs from traditional CAE workflows, why inverse design matters, and how Flexcompute is making advanced optimization tools accessible across disciplines.
Can you start by introducing yourselves, your work at Flexcompute and MIT, along with an overview of what you’ll be presenting at CDFAM Amsterdam?
Momchil: At Flexcompute, I am leading the development of our Tidy3D solver, our flagship tool for industries like integrated photonics, communications, quantum computing devices, etc.
Tidy3D is powering our photonic design automation (PDA) platform, a comprehensive suite of tools for designing and optimizing integrated photonic devices, including layout tool and an advanced inverse design functionality.
Qiqi: At Flexcompute, I focus on Geometry Enriched Multiphysics, a bridge between parametric CAD and multi-physics, multi-fidelity physics simulation and AI. My MIT work spans numerical algorithms and complex dynamical systems.
Your talk focuses on Computer-Aided Optimization (CAO) and how GPU-native simulations are accelerating this shift. How would you describe the fundamental differences between traditional CAE workflows and the CAO approach you’re helping to develop?
Traditional CAE involves slow, manual “design-simulate-analyze-redesign” loops, where simulation is often a bottleneck. Optimization is typically separate and limited.
Our CAO approach, powered by Flexcompute’s GPU-native speed, tightly integrates simulation within automated optimization loops.
This allows for exploring vast design spaces quickly, shifting engineers from manual iteration to defining optimization problems and leveraging AI-driven design discovery.
For engineers working with traditional simulation and design workflows, how would they know it’s time to start exploring tools like yours? What signs or challenges in their current process usually indicate that a shift to CAO and GPU-native simulation could make a real difference?
It’s time to explore our tools if you face:
Simulation bottlenecks that are limiting design iterations and time-to-market.
Difficulty exploring complex design spaces or achieving significant performance gains with traditional methods.
A need for high-fidelity results that are too slow with current CPU-based solvers.
A desire to automate design optimization and leverage techniques like inverse design.
You integrate techniques like adjoint optimization, inverse design, and sensitivity analysis into your workflow. How do you choose which optimization strategy to apply depending on the complexity or objectives of a given design problem?
Various considerations go in picking the right strategy / tool for a given problem:
Sensitivity Analysis: This is typically used first to understand which design parameters most significantly impact performance, helping to focus optimization efforts or simplify the design space.
Adjoint Methods for Inverse Design & Gradient-Based Optimization: When we have a clear performance target or need to optimize designs with many variables (like complex shapes in aerodynamics or photonics), adjoint methods are exceptionally powerful.
They efficiently compute design sensitivities (gradients), enabling us to ‘invert’ the design problem to find the optimal parameters or to directly optimize towards an objective.
Global Optimization & ML-Inspired Approaches (e.g., Bayesian Optimization, Particle Swarm): For design spaces that are highly complex, non-convex, or where gradient information is hard to obtain or less reliable, global search strategies or machine learning-based optimization can be very effective.
These are particularly useful when the number of design parameters is moderate, and we need to explore broadly to avoid local minima and find truly novel solutions.
At Flexcompute, we automate all of these to the extent that running an entire optimization is as simple as running a single simulation.
High-fidelity GPU-native simulation promises to dramatically speed up iteration. How has this access to faster simulation changed the types of design problems that are now practical to explore?
Speed and scale go hand in hand. We often talk about shortening turnaround time for a specific simulation, but in fact what is just as relevant is the vastly larger amount of compute that customers can now complete within the same time. This allows them to
Tackle larger, more complex problems with higher fidelity.
Extensively search the parameter space to find both optimal and robust (e.g., considering uncertainties) designs.
Produce more innovative outcomes through rapid what-if scenarios and the use of computationally intensive methods like generative design.
Your applications span very different fields, from aerospace aerodynamics to photonic devices. What principles or challenges do you see as consistent across these domains when applying CAO techniques?
The universal need for fast, accurate simulations to drive optimization.
The power of numerical optimization methods across different physics.
The drive to automate the design loop for better, faster results.
Consistent Challenges Are:
Clearly defining the optimization problem (objectives, constraints).
Handling high-dimensional design spaces and multi-objective problems.
Integrating CAO into existing workflows and ensuring manufacturability.
As you present at CDFAM Amsterdam, what are you most excited to share with the broader computational design audience, and what kinds of conversations or collaborations are you hoping to build?
Momchil: I’m particularly excited to present our “inverse design for everyone” photonics platform.
Inverse design has long been popular in academia, and it was part of my research interests already back at Stanford. However, due to a steep learning curve, and fabricability concerns, it has not yet been embraced by industry.
Flexcompute is now driving that adoption by removing entry barriers, implementing advanced techniques to ensure fabricability, and demonstrating automated optimization cycles producing best-in-class devices.
We hope to connect with industry leaders, academic researchers, and software developers to discuss future challenges, explore new applications for our technology, and foster collaborations to push the boundaries of computational design.
To learn more about how GPU-native simulation and inverse design are advancing fields from aerospace to integrated photonics, don’t miss the opportunity to connect with Momchil Minkov, Qiqi Wang, and others like them at CDFAM Amsterdam 2025.
Flexcompute’s site includes extensive educational resources on simulation and optimization, and they’re currently hiring across multiple roles.
Register to attend and be part of the conversations shaping the next generation of computational design.