Yang Deng

My path into machine learning went through physics. I started in optical engineering — lenses, light, the mathematics of how electromagnetic fields behave in structured media. Ph.D. at Duke under Willie J. Padilla was where it all collapsed into one question: when the design space is enormous and the simulations are expensive, how do you find good designs fast?

That question turned out to be bigger than metamaterials. Surrogates, inverse design, Bayesian optimization, active learning — each a different way of stretching a simulation budget further. My thesis pushed a five-orders-of-magnitude speedup on one benchmark. My best papers are the ones that say clearly what would generalize and what wouldn't.

At an early-stage AI startup I built the infrastructure for that at production scale — connecting physics simulations to AI interfaces, writing the benchmarks that kept the agents honest. I joined as the 2nd engineer because the chance to shape something from near-zero matters more to me than most things.

Now I'm an AI Research Scientist at Axiomatic AI in Cambridge, working on verifiable AI for photonics and beyond in science — models whose answers can be checked against the physics, not just trusted.