MLIPX: machine-learned interatomic potential eXploration

The rapid advancement in machine-learned interatomic potentials (MLIPs) and the proliferation of universal MLIPs (<i>u</i>MLIPs) have significantly broadened their application scope. Community benchmarks and leaderboard rankings are frequently updated, providing statistical insights into overall progress. However, the allure of using top-performing<i>u</i>MLIPs from these leaderboards blindly to real-world applications can result in unreliable predictions if their limitations and caveats are not well understood. Fine-tuning an<i>u</i>MLIP or constructing an MLIP based on active learning are often necessary to get reasonable predictions on real-world datasets. The MLIP eXploration (MLIPX) ecosystem adopts a user-centric perspective to address the question:<i>Among the given list of MLIPs, which one should I choose for my specific application?</i>and re-evaluate it seamlessly as soon as a new MLIP arrives.MLIPXachieves this through a framework of reusable recipes for a variety of simulation tasks, automated data versioning, and integrated comparative visualization tools. This significantly reduces the overhead of setting up and analyzing results from multiple MLIPs. We present example application cases to compare different leading<i>u</i>MLIPs, showcasing the utility ofMLIPX. TheMLIPXsoftware enables users to build and share recipes for application-specific test sets, featuring powerful and interactive comparison tools via the ZnDraw web interface. Furthermore, we introduce theMLIPX-hub, fostering community engagement for the continuous development of new test cases. Our systematic framework,MLIPX, offers a reproducible and reusable solution with a rich comparison and visualization ecosystem, addressing the need for comprehensive tools to evaluate MLIPs effectively.