Neural Networks from first-principles for rich datatypes

We design neural networks from first-princples to rigorously represent the rich datatypes present in the physical sciences. For example, we’ve created (3D) Euclidean neural networks and the accompanying software package e3nn, to naturally handle as geometry and geometric tensors (scalars, vectors, matrices), which transform predicably under 3D rotations and translation; these are the datatypes of physical systems in 3D.

Accelerating existing techniques and creating new capabilities for computational chemistry and material science

The “holy grail” of computational chemistry and materials discovery is to create an algorithm such that given a list of desirable properties, the algorithm would return an arrangement of atoms with those properties. We’re helping get us there by

• generating better starting points for expensive quantum mechanical calculations so they can converge faster,
• creating surrogate models to emulate these calculations altogether,
• designing algorithms to propose new hypothetical atomic structures that we can then study with existing methods, and
• aiding in the structural characterization of new materials.