AI for Science

My work in AI for Science centers on developing machine learning architectures that can navigate the vast, high-dimensional design spaces inherent to physical systems. Traditional simulation-driven approaches to device design — particularly in nanophotonics — are computationally prohibitive when searching over continuous parameter spaces with complex physical constraints.

To address this, I have developed and led teams building generative models (autoencoders, GANs, diffusion models, and annealing-based samplers) that learn compressed representations of physically valid designs. Our Machine Learning Framework for Quantum Sampling (Applied Physics Reviews, 2021) introduced adversarial training on quantum hardware to sample from highly-constrained optimization landscapes, resulting in a 16,200% speed-up over conventional methods and becoming a major milestone for the DOE Quantum Science Center.

More recently, PearSAN (Advanced Optical Materials, 2026) introduced Pearson-correlated surrogate annealing for inverse design, enabling optimization in learned latent spaces without differentiable simulators. Our review in npj Nanophotonics (2026) synthesizes the field's progress and identifies open challenges in bridging data-driven and physics-informed paradigms.

On the application side, our attention-based authentication system (Advanced Photonics, 2024) tackled the $75B counterfeit semiconductor market by detecting tampered optical responses with a 55,000x speed-up and 41% accuracy improvement. At Quantinuum, I led development of transformer architectures (GPT-2 with constrained decoders and custom PPO/GRPO losses) for generating quantum chemistry circuits in collaboration with NVIDIA and DeepMind.

Across these efforts, the unifying theme is building ML systems that respect physical law while dramatically accelerating the path from concept to validated design.