Multi-frequency Finite Difference Frequency Domain electromagnetic solver with automatic differentiation capability provided using HIPS autograd. This type of solver is particularly useful for time-dependent photonic systems with vastly different timescales involved, e.g., an electro-optic modulation (<10 GHz) of a telecom wave (100s of THz). This autodiff-compatible solver is available under the Fan Group's finite difference electromagnetics tool ceviche.

Links: GitHub

Radiative heat transfer between time-dependent photonic systems

I extended the Fan Group's photonic computational tool called MESH to compute radiative heat transfer between objects with arbitrary spatiotemporal periodicity in both the near-field and far-field. This tool is based on Rigorous Coupled-Wave Analysis and was used in our recent paper in Physical Review Letters linked below.

Links: Code, Paper and Popular Press

Learning closed form math from data

In this project for CS 221, we developed a deep-learning algorithm for symbolic regression, where we trained a sequence-like and tree-like LSTM encoder/decoder system to produce best-fit mathematical expressions from the function data in the form of (x, y) pairs.

Links: Code, Report and Poster

Spatio-temporal deep neural networks for video classification

In this project for CS 230, we developed a deep-learning algorithm to predict one of six types of tennis shots from an RGB video input by training on the THETIS dataset.

Links: Code, Report and Poster