dpdata: A Scalable Python Toolkit for Atomistic Machine Learning Data Sets
Jinzhe Zeng, Xulong Peng, Yong‐Bin Zhuang, Haidi Wang, Haidi Wang, Fengbo Yuan, Duo Zhang, Renxi Liu, Yingze Wang, Ping Tuo, Yuzhi Zhang, Yixiao Chen, Yifan Li, Cao Thang Nguyen, Jiameng Huang, Anyang Peng, Marián Rynik, Weihong Xu, Zezhong Zhang, Xuyuan Zhou, Tao Chen, Jiahao Fan, Wanrun Jiang, Bowen Li, Denan Li, Haoxi Li, Wenshuo Liang, Ruihao Liao, Liping Liu, Chenxing Luo, Logan Ward, Kaiwei Wan, Junjie Wang, Pan Xiang, Chengqian Zhang, Jinchao Zhang, R. Zhou, Jiaxin Zhu, Linfeng Zhang, Han Wang, Han Wang
Abstract
Seamless management of atomistic data sets is a critical prerequisite for the successful development and deployment of machine learning potentials (MLPs). Here, we present dpdata, an open-source Python library designed to streamline every aspect of MLP data handling. Built upon a flexible, plugin-based architecture, dpdata supports reading, writing, and converting between a broad range of file formats─from popular quantum-chemistry packages and molecular-dynamics engines to specialized MLP frameworks. Users may define custom data types, formats, drivers, and minimizers, enabling effortless extension to emerging software. Key utilities include automated train-test splitting, coordinate perturbation for active learning, outlier-energy removal, Δ-learning data set generation, error-metric computation, and unit conversion. Through efficient NumPy-backed storage and system-level operations, dpdata achieves significant memory saving and inference speedups over configuration-by-configuration tools such as ASE. We also highlight practical impact, with dpdata used across published studies, for format conversion, data storage, coordinate perturbation, and utilization in other projects for data processing.