Applications of Automatic Differentiation to Materials Design

Developments in automatic differentiation (AD) have opened the door to an array of new possibilities throughout the physical sciences. However, AD in materials design has largely been out of reach. Simulations of novel materials were far too computationally intensive for a tool that requires running hundreds of thousands of simulations. Recent work has begun to integrate molecular dynamics simulations with AD, bringing the developments in AD to a new and exciting domain. We demonstrate the power of AD in materials design by building on a seminal paper by Torquato. The prior work begins with a specific model pair potential, and varies four parameters to achieve a honeycomb lattice. Through AD, we are able to start from an entirely arbitrary pair potential, optimize that potential, and ultimately reach a honeycomb lattice with fewer defects.