Accelerating Physics-based Models with Generative Neural Networks
Over the past few years, the potential of machine learning methods that can generate text and images has expanded significantly. This project seeks to explore the implications of these advanced techniques for simulating and characterizing chemical and biological systems computationally. While physics-based models of biomolecular systems have been studied extensively, the computational burden of conducting simulations can be prohibitively high, hindering progress on important questions, e.g. small molecule drug design. The so-called "generative models" can potentially reduce costs associated with modeling realistic physical dynamics of molecular systems by learning how to directly generate relevant system configurations. Our objective is to investigate how generative models commonly used in traditional "AI" tasks can be adapted to generate chemical and biological systems, incorporating physical intuition such as symmetry and volumetric constraints. Additionally, we will focus on developing computational strategies that leverage these powerful generative models to accelerate computations while, most importantly, ensuring reliable results.