A Tutorial Introduction to developing and deploying machine learning interatomic potentials using KLIFF

Machine learning interatomic potentials (MLIP) are now increasingly being used in materials simulations. They provide near first-principle’s accuracy at a computational cost comparable to empirical force fields. MLIPs have been growing in complexity and accuracy, however most are developed and used by a handful of involved researchers, with no easy means to transfer them between simulation codes or even research groups. These limitations have been addressed in recent extensions to the KIM-based Learning-Integrated Fitting Framework (KLIFF) package, including direct access to training data in the ColabFit repository, integrated uncertainty quantification tools, and the development of a portable MLIP implementation conforming to the KIM application programming interface (API) standard (MLIP Model Driver). The new Model Driver allows MLIPs to be deployed across multiple codes that support the KIM API (such as LAMMPS, ASE, DL_POLY, GULP, and others) with full support for HPC parallelism. In addition, the Model Driver includes an automatic differentiated high-performance descriptor library, libdescriptor, which is compatible with the PyTorch Autograd API, and can be used to transfer descriptors from a Python training environment to a C++ production environment. This tutorial will describe these extensions to KLIFF and demonstrate how to use them. All necessary tools and code samples used in the tutorial will be provided in a downloadable “KIM Development Platform” Docker image so that interested participants can follow along on their own laptops.