Evidential Deep Learning for Interatomic Potential

Machine Learning Interatomic Potentials (MLIPs) are models that utilize machine learning techniques to fit interatomic potential functions, with training data derived from ab initio methods. Consequently, MLIPs can achieve ab initio potential function accuracy with significantly faster inference times and reduced computational resource consumption. However, the datasets required for training MLIPs at ab initio accuracy are inherently resource-intensive and cannot encompass all possible configurations. When MLIP models trained on these datasets are employed in molecular dynamics (MD) simulations, they may encounter out-of-distribution (OOD) data, leading to a collapse of the MD simulation. To mitigate this issue, active learning approaches can be employed, iteratively sampling OOD data to enrich the training database. Nonetheless, conventional methods often require substantial time or result in decreased MLIP model accuracy. We propose a novel uncertainty output method that effectively balances speed and accuracy, demonstrating excellent performance.The dataset is for the article "Evidential Deep Learning for Interatomic Potential",The dataset "liquid_aimd.pt" was used as the OOD data in water dataset, the dataset "newliquid_shifted" was used as the ID dataThe folder "small_molecule_weight" and "water_dataset_weight" contains the model weight documents of experiments performed in main textThe folder "UDD-MD" contains the model weight used to reproduce the result in Fig 4 in main text