BLIPs: Bayesian Learned Interatomic Potentials

Machine Learning Interatomic Potentials (MLIPs) are becoming a central tool in simulation-based chemistry. However, like most deep learning models, MLIPs struggle to make accurate predictions on out-of-distribution data or when trained in a data-scarce regime, both common scenarios in simulation-based chemistry. Moreover, MLIPs do not provide uncertainty estimates by construction, which are fundamental to guide active learning pipelines and to ensure the accuracy of simulation results compared to quantum calculations. To address this shortcoming, we propose BLIPs: Bayesian Learned Interatomic Potentials. BLIP is a scalable, architecture-agnostic variational Bayesian framework for training or fine-tuning MLIPs, built on an adaptive version of Variational Dropout. BLIP delivers well-calibrated uncertainty estimates and minimal computational overhead for energy and forces prediction at inference time, while integrating seamlessly with (equivariant) message-passing architectures. Empirical results on simulation-based computational chemistry tasks demonstrate improved predictive accuracy with respect to standard MLIPs, and trustworthy uncertainty estimates, especially in data-scarse or heavy out-of-distribution regimes. Moreover, fine-tuning pretrained MLIPs with BLIP yields consistent performance gains and calibrated uncertainties.

Simulating how atoms and molecules behave is a powerful way to discover new materials and drugs, but it requires enormous computational resources when performed with full quantum-mechanical accuracy. AI-based models called interatomic potentials have emerged as a faster alternative, learning to mimic quantum calculations from data. However, like most AI systems, these models can fail silently when asked to make predictions in situations they have not seen before, and they give no warning when they are likely to be wrong. We introduce BLIPs, a framework that wraps existing interatomic potential models in a principled statistical approach called Bayesian inference. Instead of a single prediction, BLIPs produces a range of predictions reflecting how confident the model is, at almost no extra computational cost. In practice, BLIPs makes models more accurate in data-limited and unfamiliar scenarios, produces reliable uncertainty estimates to guide automated data collection, and works as a plug-in to improve already-trained models.