Evidential Deep Learning for Interatomic Potentials

Dataset DescriptionThis repository provides model checkpoints, simulation trajectories, and implementation code accompanying the paper “Evidential Deep Learning for Interatomic Potentials.”ckpt.zip: Contains checkpoints of the eIP model trained on multiple datasets. For the silica glass dataset, additional checkpoints are included for models trained with ensemble, Monte Carlo dropout, Gaussian Mixture Model (GMM), and Maximum Variance Estimation (MVE).traj.zip: Includes molecular dynamics (MD) and uncertainty-driven dynamics (UDD) trajectories for water, lithium iron phosphate (LiFePO₄), and polydimethylsiloxane (PDMS). All trajectories are stored in the extended XYZ format.AlphaNet&TorchMDNet.zip: Contains code implementations based on AlphaNet and TorchMD-Net backbones, integrating uncertainty quantification (UQ) methods including eIP and DPOSE (shallow ensemble).Source Data.zip: Contains the source data for paper “Evidential Deep Learning for Interatomic Potentials.”

数据集说明
本仓库配套于论文《原子间势的证据式深度学习》（"Evidential Deep Learning for Interatomic Potentials"），提供模型检查点（checkpoint）、模拟轨迹与实现代码。
ckpt.zip：内含基于多数据集训练得到的eIP模型的检查点。针对二氧化硅玻璃数据集，额外提供了采用集成学习、蒙特卡洛丢弃（Monte Carlo dropout）、高斯混合模型（Gaussian Mixture Model, GMM）以及最大方差估计（Maximum Variance Estimation, MVE）方法训练的模型的检查点。
traj.zip：内含针对水、磷酸铁锂（LiFePO₄）以及聚二甲基硅氧烷（PDMS）的分子动力学（molecular dynamics, MD）轨迹与不确定性驱动动力学（uncertainty-driven dynamics, UDD）轨迹。所有轨迹均采用扩展XYZ格式存储。
AlphaNet&TorchMDNet.zip：内含基于AlphaNet与TorchMD-Net骨干网络的代码实现，集成了包括eIP与DPOSE（浅层集成）在内的不确定性量化（uncertainty quantification, UQ）方法。
Source Data.zip：内含论文《原子间势的证据式深度学习》（"Evidential Deep Learning for Interatomic Potentials"）的原始数据集。