Machine learning potentials (MLPs) have emerged as powerful simulation tools for heterogeneous catalysis. While current MD-active-learning workflows excel at fitting a specific system/reaction pathway through iterative structural sampling, the transition toward more general and transferable MLPs, designed to handle diverse structures and reactive events within a defined chemical space, presents fundamentally new challenges. Such generality often requires highly diverse, nonequilibrium training data, for which standard practices may confront challenges regarding training discipline and evaluation logic. Here, using our recently developed REICO method as an example to generate such data sets, we systematically investigate the distinct pathologies that arise when training on such diverse data, revealing critical deviations from standard system-specific MLP training. We further provide detailed recommendations on data cleaning, model selection, and error metrics for both numerical performance and physical validation, offering practical guidance for training MLPs with diverse and hybrid data sets.
Wu et al. (2026) studied this question.