P-DRUM: Post-hoc Descriptor-based Residual Uncertainty Modeling for Machine Learning Potentials

P-DRUM, or Post-hoc Descriptor-based Residual Uncertainty Modeling, presents a significant advancement in the field of uncertainty quantification (UQ) for machine learning interatomic potentials (MLIPs). Traditional ensemble methods are regarded as the gold standard for UQ but are often hindered by high computational costs. P-DRUM offers a simpler and more efficient alternative by utilizing the descriptors of trained graph neural networks to estimate residual errors, which serve as proxies for prediction uncertainty. This innovative approach not only enhances computational efficiency but also maintains accuracy, as it models the discrepancies between MLIP predictions and ground truth values. The paper benchmarks P-DRUM against established UQ methods, exploring its effectiveness and limitations, thereby contributing valuable insights to the ongoing discourse on improving UQ in machine learning applications.