Interpreting neural networks: Physical insight from supervised committor classifiers
Date:
Contributed talk at the 2025 CoSeC Conference.
CoSeC is the Computational Science Centre for Research Communities.
Abstract
A central challenge in chemical physics is identifying low-dimensional descriptors of rare, complex processes. Such descriptors, also known as reaction coordinates, are integral to several classic rate theories, and various methods have been explored to obtain them[1,2], including machine learning approaches [3,4]. The committor (pB), the probability that a system reaches state B before returning to state A, is the optimal one-dimensional descriptor for rare transitions between long-lived states [5]. We develop a tailored neural network [6] to learn the committor from GPU-accelerated simulations [7] in a standard model system (namely the 2D Ising model), enabling reproducible and scalable computation. Using activation maps, we analyze the network to uncover key features missed by classical theories, with the goal of capturing these details in an improved non-machine-learned descriptor. This approach provides interpretable insights that may extend beyond the specific system studied.
References
[1] B. Peters, Annu. Rev. Phys. Chem. 67, 669-690 (2016).
[2] B. Peters and B. L. Trout, Annu. Rev. Phys. Chem. 67, 669-690 (2016).
[3] D. Aristoff, et al., J. Chem. Phys. 161, 084113 (2024).
[4] Ma, A., J. Phys. Chem. B 109, 6769-6779 (2005).
[5] Krivov, J. Chem. Theory Comput. 9, 135 (2013).
[6] H. J. Naguszewski, GitHub, 2025, https://github.com/HubertJN/committor-predictor
[7] D. Quigley, H. J. Naguszewski, GitHub, 2025, https://github.com/dquigley533/GASP