Scalable Bayesian Uncertainty Quantification for Neural Network Potentials: Promise and Pitfalls.

Area of Science:

• Computational chemistry
• Machine learning in materials science
• Statistical physics

Background:

• Neural network (NN) potentials offer computational efficiency for molecular dynamics (MD) simulations.
• NN potentials can produce inaccurate predictions when used outside their training data domains.
• Accurate uncertainty quantification (UQ) is essential for reliable NN potential-based MD simulations.

Purpose of the Study:

• To demonstrate the efficacy of scalable Bayesian UQ for NN potentials using stochastic gradient MCMC (SG-MCMC).
• To assess the impact of training data size and multiple Markov chains on UQ reliability.
• To compare SG-MCMC with the Deep Ensemble method for UQ in MD.

Main Methods:

• Training graph NN potentials for coarse-grained liquid water and alanine dipeptide systems.
• Applying SG-MCMC for Bayesian UQ of NN potentials.
• Investigating the effect of cold posteriors on training data requirements.
• Utilizing multiple Markov chains for robust UQ.
• Comparing SG-MCMC with the Deep Ensemble method.

Main Results:

• SG-MCMC provides reliable uncertainty estimates for MD observables.
• Cold posteriors can decrease the necessary training data size for UQ.
• Multiple Markov chains are required for dependable UQ.
• SG-MCMC and Deep Ensemble methods show comparable performance in capturing aleatoric and epistemic uncertainty.
• Neither method reliably captures systematic uncertainty.

Conclusions:

• Scalable Bayesian UQ via SG-MCMC is a viable approach for NN potentials in MD.
• Careful modeling is needed to minimize systematic uncertainty for accurate credible intervals.
• This work advances trustworthy NN potential-based MD simulations for practical decision-making.