KUSP: Python server for deploying ML interatomic potentials

Amit Gupta, Ellad B. Tadmor, Stefano Martiniani

Submission Track: Full Paper

Submission Category: AI-Guided Design

Keywords: machine-learned interatomic potentials, KIM API, openkim, model validation, model development

Abstract: The KIM Utility for Serving Potentials (KUSP) is a Python package designed to facilitate the rapid deployment of machine-learned interatomic potentials (MLIPs) to arbitrary simulation codes. It imposes minimal restrictions on ML architecture or libraries and is compatible with popular ML frameworks such as PyTorch, JAX, and TensorFlow, as well as utilities like PyTorch Geometric and the Deep Graph Library. By providing a simulator-agnostic interface via the KIM API, KUSP allows researchers to quickly prototype and benchmark their models against OpenKIM "tests" (molecular simulation-based material property calculations) and "verification checks" (basic physical consistency checks), ensuring that valid ML models are on equal footing with OpenKIM-supported interatomic potentials. KUSP employs a client-server architecture where the Python server communicates with the KIM API using sockets, converting model output to a KIM API-compatible format. This tool is particularly valuable for model validation and development, offering insights that can guide improvements in models and datasets.

Submission Number: 13