Supervised Learning and the Finite-Temperature String Method for Computing Committor Functions and Reaction Rates

Abstract

A central object in the computational studies of rare events is the committor function. Though costly to compute, the committor function encodes complete mechanistic information of the processes involving rare events, including reaction rates and transition-state ensembles. Under the framework of transition path theory (TPT), recent work [1] proposes an algorithm where a feedback loop couples a neural network that models the committor function with importance sampling, mainly umbrella sampling, which collects data needed for adaptive training. In this work, we show additional modifications are needed to improve the accuracy of the algorithm. The first modification adds elements of supervised learning, which allows the neural network to improve its prediction by fitting to sample-mean estimates of committor values obtained from short molecular dynamics trajectories. The second modification replaces the committor-based umbrella sampling with the finite-temperature string (FTS) method, which enables homogeneous sampling in regions where transition pathways are located. We test our modifications on low-dimensional systems with non-convex potential energy where reference solutions can be found via analytical or the finite element methods, and show how combining supervised learning and the FTS method yields accurate computation of committor functions and reaction rates. We also provide an error analysis for algorithms that use the FTS method, using which reaction rates can be accurately estimated during training with a small number of samples. The methods are then applied to a molecular system in which no reference solution is known, where accurate computations of committor functions and reaction rates can still be obtained.

arXiv comment39 pages main text, 17 pages appendix, 41 figures, one table

References in corpus (16)

• Adam: A Method for Stochastic Optimization
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Escaping free-energy minima
• SchNet - a deep learning architecture for molecules and materials
• E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials
• Statistically optimal analysis of samples from multiple equilibrium states
• Fast Graph Representation Learning with PyTorch Geometric
• Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds
• Gradient Descent Provably Optimizes Over-parameterized Neural Networks
• Gradient Descent Finds Global Minima of Deep Neural Networks
• Computing Committor Functions for the Study of Rare Events Using Deep Learning
• Eigenvector method for umbrella sampling enables error analysis
• A Theory of Localized Excitations in Supercooled Liquids
• Active Importance Sampling for Variational Objectives Dominated by Rare Events: Consequences for Optimization and Generalization
• A semigroup method for high dimensional committor functions based on neural network
• Patterns, predictions, and actions: A story about machine learning

Cited by in corpus (3)

• Variational deep learning of equilibrium transition path ensembles
• From high-dimensional committors to reactive insights
• Deep learning of committor and explainable artificial intelligence analysis for identifying reaction coordinates