Scalable Bayesian Uncertainty Quantification for Neural Network Potentials: Promise and Pitfalls
This work addresses the need for trustworthy uncertainty quantification in neural network potential-based molecular dynamics simulations, which is crucial for practical decision-making in fields like chemistry and materials science, though it is incremental as it builds on existing Bayesian and ensemble methods.
The paper tackles the problem of unreliable predictions from neural network potentials in molecular dynamics simulations when applied outside their training domain, demonstrating that scalable Bayesian uncertainty quantification via stochastic gradient MCMC yields reliable uncertainty estimates for observables, with methods achieving comparable results to Deep Ensembles but requiring less hyperparameter tuning.
Neural network (NN) potentials promise highly accurate molecular dynamics (MD) simulations within the computational complexity of classical MD force fields. However, when applied outside their training domain, NN potential predictions can be inaccurate, increasing the need for Uncertainty Quantification (UQ). Bayesian modeling provides the mathematical framework for UQ, but classical Bayesian methods based on Markov chain Monte Carlo (MCMC) are computationally intractable for NN potentials. By training graph NN potentials for coarse-grained systems of liquid water and alanine dipeptide, we demonstrate here that scalable Bayesian UQ via stochastic gradient MCMC (SG-MCMC) yields reliable uncertainty estimates for MD observables. We show that cold posteriors can reduce the required training data size and that for reliable UQ, multiple Markov chains are needed. Additionally, we find that SG-MCMC and the Deep Ensemble method achieve comparable results, despite shorter training and less hyperparameter tuning of the latter. We show that both methods can capture aleatoric and epistemic uncertainty reliably, but not systematic uncertainty, which needs to be minimized by adequate modeling to obtain accurate credible intervals for MD observables. Our results represent a step towards accurate UQ that is of vital importance for trustworthy NN potential-based MD simulations required for decision-making in practice.