Orchestrating Interatomic Potential Training and Analysis

Machine learned interatomic potentials have been a boon for enabling large-scale molecular dynamics simulations near ab initio accuracies but at a fraction of the cost. Yet despite these strides forward, significant challenges remain. To generate a viable potential, one needs to curate an appropriate training data set, select and tune the potential form, and execute an effective training procedure. Once an accurate model is obtained, it can be deployed in subsequent calculations to predict material properties. Each of these steps requires the management of disparate computational resources alongside analysis to quantify uncertainties that arise throughout the workflow – from training to property prediction. We have developed a tool to coordinate these tasks with minimal human intervention and incorporate advanced quality metrics during the training of potentials to effectively curate our training data and better estimate the uncertainty of predicted properties. By introducing a systematic treatment of the training and analysis procedures, we aim to prevent unintentional bias while simultaneously training more efficiently. These advances are key to leveraging interatomic potentials for characterizing diverse materials across a range of applications and conditions.