Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Atomistic simulations

Background:

• Neural network (NN) interatomic potentials offer fast predictions but struggle with extrapolation beyond their training data.
• Uncertainty quantification identifies unreliable predictions but requires extensive sampling, often via computationally expensive molecular dynamics simulations.
• Current methods for improving NN potentials are limited by the need for exhaustive phase space exploration.

Purpose of the Study:

• To develop an efficient method for sampling high-uncertainty configurations to improve neural network potentials.
• To reduce the computational cost associated with training and refining NN interatomic potentials.
• To enhance the reliability and accuracy of NN potentials for predicting material properties.

Main Methods:

• Exploiting automatic differentiation to guide atomistic systems toward high-uncertainty configurations.
• Employing adversarial attacks on uncertainty metrics to identify informative geometries for NN training domain expansion.
• Integrating the sampling strategy with an active learning loop to iteratively improve NN potentials.

Main Results:

• Demonstrated efficient sampling of kinetic barriers, molecular collective variables, and zeolite-molecule interactions.
• Successfully bootstrapped and improved NN potentials with fewer calls to ground truth electronic structure methods.
• Showcased the method's applicability across diverse molecular and materials systems and NN architectures.

Conclusions:

• The developed approach significantly enhances the efficiency and effectiveness of training and refining neural network interatomic potentials.
• This method offers a powerful tool for expanding the predictive domain of NN potentials, crucial for accurate materials modeling.
• The technique is broadly applicable, promising advancements in computational chemistry and materials science research.