Interrogating the neural network black box: Physical insight from supervised-encoded probability proxies
Date:
Contributed talk at the 2025 7th HPC-AI Advisory Council.
Abstract
A central challenge in chemical physics is identifying low-dimensional descriptors of rare, complex processes. The committor (pB) , the probability a system reaches state B before returning to state A, is the optimal one-dimensional descriptor for rare transitions between long-lived states [1]. We develop a tailored neural network to learn the committor from GPU-accelerated simulations in a standard model system. Using class activation maps, we analyze the network to uncover key features missed by classical theories. This approach allows us to validate the model’s predictive power through physically meaningful variables typically overlooked in traditional descriptions of such rare events.
References
[1] Krivov, On reaction coordinate optimality, J. Chem. Theory Comput. 9, 135 (2013).