Training, validation, and interpretation of committor models based on reweighted path ensembles.

The committor function is often viewed as the optimal reaction coordinate for rare-event dynamics and has become increasingly accessible through advances in machine learning and path sampling methods. The recent introduction of Artificial Intelligence Molecular Mechanism Discovery combined with Transition Interface Sampling (AIMMD-TIS) enables the construction of neural net committor models across the full transition region. However, comparatively little attention has been devoted to understanding why such models can be learned, how they should be diagnosed and validated, and how mechanistic insights can be extracted from them. Here, we provide concepts and tools to obtain statistical intuition, diagnostics, validation, and interpretation of committor models trained on reweighted path ensembles. We illustrate these concepts and tools on a high-barrier host-guest (un)binding process using AIMMD-TIS and MultiSetMBAR reweighting. Together, our methodology provides a framework for understanding, validating, and interpreting learned committor models.