[Paper Review] On-the-Fly Active Learning of Interpretable Bayesian Force Fields for Atomistic Rare Events

Jonathan Vandermause, Steven B. Torrisi+5 more|arXiv (Cornell University)|Apr 3, 2019

Machine Learning in Materials Science4 citations

TL;DR

This paper introduces FLARE, a novel on-the-fly active learning framework that uses interpretable, low-dimensional Bayesian force fields based on Gaussian process regression to model atomistic rare events with minimal ab initio training data. By leveraging model uncertainty to guide dynamic data acquisition during molecular dynamics, the method achieves high accuracy with ~100 DFT calculations while enabling automatic, uncertainty-driven model updates for complex systems including multi-element materials and rare diffusive events.

ABSTRACT

Machine learned force fields typically require manual construction of training sets consisting of thousands of first principles calculations, which can result in low training efficiency and unpredictable errors when applied to structures not represented in the training set of the model. This severely limits the practical application of these models in systems with dynamics governed by important rare events, such as chemical reactions and diffusion. We present an adaptive Bayesian inference method for automating the training of interpretable, low-dimensional, and multi-element interatomic force fields using structures drawn on the fly from molecular dynamics simulations. Within an active learning framework, the internal uncertainty of a Gaussian process regression model is used to decide whether to accept the model prediction or to perform a first principles calculation to augment the training set of the model. The method is applied to a range of single- and multi-element systems and shown to achieve a favorable balance of accuracy and computational efficiency, while requiring a minimal amount of ab initio training data. We provide a fully open-source implementation of our method, as well as a procedure to map trained models to computationally efficient tabulated force fields.

Motivation & Objective

• To address the high computational cost and poor generalization of traditional machine-learned force fields that require extensive manual training data.
• To develop a method that automatically identifies and computes critical DFT data points during molecular dynamics simulations, reducing reliance on pre-specified training sets.
• To provide trustworthy, interpretable uncertainty estimates in force field predictions to guide model improvement and ensure reliability in rare event simulations.
• To enable efficient, scalable, and automated training of force fields for complex materials with minimal human intervention.
• To map trained Gaussian process models to computationally efficient tabulated spline force fields for production-scale simulations.

Proposed method

• The method employs a Gaussian process regression model with a physically motivated, low-dimensional n-body kernel (2- and 3-body) to model interatomic energies and forces, ensuring interpretability and reducing model complexity.
• Model uncertainty is quantified via the predictive variance of the GP, which drives active learning by identifying configurations where the model is most uncertain.
• During on-the-fly molecular dynamics, the algorithm decides in real time whether to trust the GP prediction or perform a new DFT calculation based on uncertainty thresholds.
• Hyperparameters of the GP are tuned in a data-driven fashion using maximum marginal likelihood estimation via BFGS optimization of the log marginal likelihood.
• The trained GP models are mapped to high-accuracy, computationally efficient cubic spline interpolants for use in large-scale MD simulations.
• The entire pipeline is implemented in the open-source FLARE software package, integrating with Quantum ESPRESSO for DFT and LAMMPS for production MD.

Experimental results

Research questions

• RQ1 Can active learning based on model uncertainty significantly reduce the number of required DFT calculations for training accurate force fields?
• RQ2 Can interpretable, low-dimensional Bayesian force fields with reliable uncertainty estimates be used to guide on-the-fly data acquisition during MD simulations?
• RQ3 Can such a framework reliably model rare events like crystal melting and diffusion in multi-element systems with minimal training data?
• RQ4 How does the performance of the resulting force field compare to traditional ML force fields in terms of accuracy and computational efficiency?
• RQ5 Can the GP-based force field be effectively mapped to a tabulated spline model without sacrificing accuracy for production-scale simulations?

Key findings

• The method achieves high accuracy with only ~100 ab initio calculations, significantly reducing the training data requirement compared to conventional approaches.
• The active learning strategy based on GP uncertainty enables reliable identification of critical configurations, such as transition states and rare events, during on-the-fly simulation.
• The resulting force fields are interpretable, as they are based on physically motivated n-body interactions, enabling direct physical insight into force contributions.
• The model demonstrates robust performance in simulating rapid crystal melts and rare diffusive events in systems like AgI and Cu-Au alloys.
• The mapping of GP models to cubic spline interpolants preserves accuracy while enabling orders-of-magnitude speedup in force evaluation for production MD.
• The open-source FLARE package enables reproducibility and broad adoption, with full support for multi-element systems and uncertainty-aware simulations.