Noise-resilient training of machine learning interatomic potentials: Terry’s first paper out in J. Chem. Phys.

Congratulations to Terry, who just published his first lead-author paper (J. Chem. Phys. 165, 144116 (2026)) over a summer project as an undergraduate student! His work is remarkably impressive, especially given that he was new to the field when he joined us.

Machine learning interatomic potentials (MLIPs) are often trained on vast datasets of reference data, but unconverged or inconsistent calculations can add unexpected and difficult-to-identify noise in the reference data, degrading the performance of the MLIP.

In this work, we introduce a noise-resilient training method that automatically down-weights noisy samples, without requiring additional reference calculations, manual filtering, or human intervention. We demonstrate that our method prevents overfitting, and matches the performance of expert-refined datasets with significantly reduced overhead; we also show that our method, when applied to a foundation MLIP for organic chemistry, increases the model’s accuracy.

Our framework thus provides a simple, automated solution for training robust models on imperfect datasets across dataset sizes.