GPU-accelerated FDTD
Re-implementing the inner simulation loop on GPUs so a generation of an inverse-design run takes seconds, not hours.
Area 03
Inverse Design for Photonic Devices
Letting algorithms design our devices — and using GPUs to do it fast.
Philosophy
Traditional photonic device design is forward-style: a human picks a geometry, simulates it (FDTD, mode solvers), tweaks parameters, and repeats. Inverse design flips this: specify the desired optical response and let an optimization algorithm — often gradient-based or ML-driven — find the geometry that achieves it. The result is unintuitive but high-performing devices that no human would have drawn.
The catch is speed. Electromagnetic simulation of a single non-trivial structure can take hours; an inverse-design loop that wants thousands of evaluations becomes weeks of compute. We accelerate the workflow with GPUs and ML surrogates, extending it from single components up to full subsystem simulation, and feeding the resulting devices back into our photonic architecture work (Areas 01, 02).
Open questions
- Can ML surrogates replace full electromagnetic simulation without losing fidelity?
- How do we propagate device-level variation up to system-level performance prediction?
- What's the right design representation for ML-driven inverse design — pixel grids, level sets, parametric primitives?
- How do we make inverse-designed devices manufacturable, not just simulatable?
Foundational reading
Where this thread started
- Molesky et al., Inverse design in nanophotonics, Nature Photonics, 2018.
- Piggott et al., Inverse design and demonstration of a compact and broadband on-chip wavelength demultiplexer, Nature Photonics, 2015.
- Recent ML-surrogate work from Stanford, MIT, and the Caltech computational nano-optics groups.
[Inverse-design pipeline diagram — pending]
Sub-themes & ongoing projects
ML surrogates for photonic simulation
Training neural networks to approximate the response of common photonic primitives, with calibrated uncertainty so we know when to fall back to full simulation.
Component-to-system inverse design
Scaling beyond individual MZIs and MRRs to inverse-designing the layout and parameters of full photonic subsystems.
Variation-aware design
Optimizing not just for nominal performance but for robust performance across realistic fabrication variation distributions.