Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials: Data

This repository includes data sets from the paper "Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials". It also provides configuration files for experiments, which can be used to reproduce the results of the paper combined with publicly available code at GitHub. Data structure for alanine dipeptide: The initial structure (ala2-ffs/ala2_init.extxyz) and the AMBER ff19SB files (ala2-ffs/input.pdb and ala2-ffs/parm7.prmtop) for reference calculations are in the respective folder. We also include the test data set generated by running molecular dynamics simulations at 1200 K using canonical (NVT) statistical ensemble (ala2-ffs/ala2_test.extxyz). Task folders like ala2-ffs/ala2-300K-ffs contain experiments under different conditions (different temperatures). For detailed setups, refer to alebrew/task_execution.py on GitHub, where each task has the same name as the respective folder. Each task includes various methods for learning potentials, with method naming convention as <presampling_method>_<params>-<simulation_method>_<params>-<bias_method>_<params>-<selection_method>_<params>-<seed>. For <bias_method>, parameters denote force biasing and relative hydrogen biasing. Compare method names with respective configuration files. Note we also include adversarial training experiments, i.e., <simulation_method> is 'adversarial', along with uncertainty-biased molecular dynamics. Data structure for MIL-53(Al): The initial structure (mil53/mil53_cp_init.extxyz) and CP2K input file (mil53/mof_base_input.txt) for reference calculations are in the respective folder. Task folders like mil53/mil53_cp-300K-0bar-v2 and mil53/mil53_cp-600K-2500bar-v1 contain experiments under different conditions (different temperatures and pressures). v1 signifies initiation from 32 randomly displaced configurations, while v2 denotes experiments using a pre-trained mode with 256 configurations from Zenodo. For detailed setups, refer to alebrew/task_execution.py on GitHub, where each task has the same name as the respective folder. Each task includes various methods for learning potentials, with method naming convention as <presampling_method>_<params>-<simulation_method>_<params>-<bias_method>_<params>-<selection_method>_<params>-<seed>. For <bias_method>, parameters denote force biasing, relative hydrogen biasing, and stress biasing strengths. Compare method names with respective configuration files.本仓库收录了论文《面向学习一致精度原子间势的不确定性偏置分子动力学（Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials）》中的数据集，同时提供了实验所需的配置文件。结合GitHub上公开的代码，即可复现该论文的研究结果。 ### 丙氨酸二肽数据集结构 参考计算所需的初始结构（ala2-ffs/ala2_init.extxyz）与AMBER ff19SB力场文件（ala2-ffs/input.pdb及ala2-ffs/parm7.prmtop）均存放于对应文件夹中。本仓库同时收录了采用正则系综（NVT）、在1200 K下运行分子动力学模拟所生成的测试数据集（ala2-ffs/ala2_test.extxyz）。诸如ala2-ffs/ala2-300K-ffs的任务文件夹包含了不同温度条件下的实验配置。如需了解详细设置，请参阅GitHub上的alebrew/task_execution.py文件，其中每个任务的名称与其对应文件夹的名称保持一致。每个任务均包含多种势函数学习方法，方法命名规则为：<预采样方法>_<参数>-<模拟方法>_<参数>-<偏置方法>_<参数>-<选择方法>_<参数>-<随机种子>。其中<偏置方法>的参数表示力偏置与相对氢偏置的相关设置。请将方法名称与对应配置文件进行比对。此外，本仓库还包含对抗训练实验，即当<模拟方法>取值为'adversarial'时，搭配不确定性偏置分子动力学方法使用。 ### MIL-53(Al)数据集结构 参考计算所需的初始结构（mil53/mil53_cp_init.extxyz）与CP2K输入文件（mil53/mof_base_input.txt）均存放于对应文件夹中。诸如mil53/mil53_cp-300K-0bar-v2和mil53/mil53_cp-600K-2500bar-v1的任务文件夹包含了不同温压条件下的实验配置。其中v1表示从32个随机位移的构型初始化，v2表示采用预训练模式、使用来自Zenodo的256个构型所开展的实验。如需了解详细设置，请参阅GitHub上的alebrew/task_execution.py文件，其中每个任务的名称与其对应文件夹的名称保持一致。每个任务均包含多种势函数学习方法，方法命名规则为：<预采样方法>_<参数>-<模拟方法>_<参数>-<偏置方法>_<参数>-<选择方法>_<参数>-<随机种子>。其中<偏置方法>的参数表示力偏置、相对氢偏置与应力偏置的强度参数。请将方法名称与对应配置文件进行比对。