1. Introduction

1.1. Overview of FitSNAP

FitSNAP is a machine learning software that focuses on modelling how interatomic interactions, or potential energy, depends on geometry between atoms. In this sense the machine learning problem is learning the mapping between atomic coordinates and potential energy. Since the physics of interatomic interactions are invariant to translation, rotation, and permutation, we must describe interatomic geometry in such a way that these symmetries are satisfied. We do this using invariant “descriptors” such as SNAP descriptors and ACE descriptors.

The resulting machine learned models therefore predict interatomic interaction energies as a function of these geometric descriptors. Once these models are obtained from FitSNAP, they may be used directly in LAMMPS to perform high-performance molecular dynamics simulations.

1.2. FitSNAP Components

Throughout these docs we will use phrases or language particular to our general way of doing molecular machine learning. These phrases are also littered throughout our code architecture, for example in the names of our Python classes. To solve molecular machine learning problems, we first need to scrape configurations of atoms from a dataset; this is achieved in the Scraper class. When storing configurations for fitting, it is useful to organize them into groups which may have different training/testing ratios or energy/force/stress weights.

Then we need a calculator to process these configurations by calculating descriptors for all atoms; this is done in the Calculator class. Once we have descriptors, we must learn the mapping between descriptors and properties like energies and forces with a solver that performs a fit via linear regression or gradient descent for neural networks; these solvers are contained in the Solver class.