KLIFF: A framework to develop physics-based and machine learning interatomic potentials

Interatomic potentials (IPs) are reduced-order models for calculating the potential energy of a system atoms given their positions in space and species. IPs treat as classical particles without explicitly modeling electrons thus computationally far less expensive than first-principles methods, enabling molecular simulations significantly larger systems over longer times. Developing an IP is complex iterative process involving multiple steps: assembling training set, designing functional form, optimizing function parameters, testing model quality, deployment to simulation packages. This paper introduces KIM-based learning-integrated fitting framework (KLIFF), package that facilitates entire development process. KLIFF supports both physics-based machine learning IPs. It adopts modular approach whereby various components process, such atomic environment descriptors, forms, loss functions, optimizers, quality analyzers, so on, work seamlessly with each other. provides flexible rapid design new forms. Trained compatible Knowledgebase Models (KIM) application programming interface (API) can be readily used major materials packages KIM, including ASE, DL_POLY, GULP, LAMMPS, QC. written Python intensive implemented C++. parallelized data shared-memory multicore desktop machines high-performance distributed memory computing clusters. We demonstrate use by Stillinger--Weber neural network silicon. The package, together its documentation, publicly available at: https://github.com/openkim/kliff.