NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials

Chengbing Chen, Yutong Li, Rui Zhao, Zhoulin Liu, Zheyong Fan, Gang Tang, Zhiyong Wang · 2025

Read NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials by Chengbing Chen, Yutong Li, Rui Zhao, Zhoulin Liu, Zheyong Fan, Gang Tang, Zhiyong Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-01868.txt.

Introduction

Chengbing Chen and co-authors release open tools that automate training sets for machine-learned atomic potentials in a 2025 paper. The neuroevolution potential method is fast and accurate, but building its training data is slow, manual work. Their NepTrain package prepares and manages data sets, and trains models on its own.

A bond-length filter weeds out unphysical structures from simulation trajectories. The companion program, NepTrainKit, gives a graphical bench for editing and exploring data. It finds outliers, picks far-apart samples, flags broken structures, and sorts configuration types. A perovskite case study walks the full workflow and validates the trained model on material properties. The authors say the pair removes a bottleneck for the whole potential-building community. Both tools are open source.

Catalog record

CREATORChengbing Chen, Yutong Li, Rui Zhao, Zhoulin Liu, Zheyong Fan, Gang Tang, Zhiyong Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-01868

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials Chengbing Chen, Yutong Li, Rui Zhao, Zhoulin Liu, Zheyong Fan, Gang Tang, Zhiyong Wang Abstract As a machine-learned potential, the neuroevolution potential (NEP) method features exceptional computational efficiency and has been successfully applied in materials science. Constructing high-quality training datasets is crucial for developing accurate NEP models. However, the preparation and screening of NEP training datasets remain a bottleneck for broader applications due to their time-consuming, labor-intensive, and resource-intensive nature. In this work, we have developed NepTrain and NepTrainKit, which are dedicated to initializing and managing training datasets to generate high-quality training sets while automating NEP model training. NepTrain is an open-source Python package that features a bond length filtering method to effectively identify and remove non-physical structures from molecular dynamics trajectories, thereby ensuring high-quality training datasets. NepTrainKit is a graphical user interface (GUI) software designed specifically for NEP training datasets, providing functionalities for data editing, visualization, and interactive exploration. It integrates key features such as outlier identification, farthest-point sampling, non-physical structure detection, and configuration type selection. The combination of these tools enables users to process datasets more efficiently and conveniently. Using CsPbI 3 \rm CsPbI_{3} as a case study, we demonstrate the complete workflow for training NEP models with NepTrain and further validate the models through materials property predictions. We believe this toolkit will greatly benefit researchers working with machine learning interatomic potentials. 1. Introduction The rapid development of machine-learned potentials (MLPs) , or machine-learned force fields Unke et al. 2021 , has provided a revolutionary solution for molecular dynamics (MD) simulations, offering both the accuracy of density functional theory (DFT) calculations and the efficiency of empirical potential. This advancement is driving the shift in computational materials science from empirical models to a data-driven paradigm. Traditional MD simulations are fundamentally dependent on the accuracy of interatomic potentials, which are typically derived from empirical formulations. However, these empirical potentials often exhibit limitations in describing complex material properties or non-equilibrium states Behler and Csányi 2021 , hindering a comprehensive capture of the associated physical phenomena. In recent years, the development of MLPs has offered a promising approach to address these challenges. Compared to conventional empirical potentials, MLP not only demonstrate higher computational accuracy but also exhibit adaptability across diverse material types, demonstrating significant potential for broad applications Friederich et al. 2021 ; Behler 2016 . Many MLP methods have been developed and widely used, including, e.g., the Behler-Parrinello neural network potential Behler and Parrinello 2007 , the Gaussian approximation potential (GAP) Bartók et al. 2010 , the spectral neighbor analysis potential Thompson et al. 2015 , the moment tensor potential (MTP) Shapeev 2016 , the deep potential Wang et al. 2018 ; Zhang et al. 2018 , the atomic cluster expansion potential Drautz 2019 , and the neuroevolution potential (NEP) Fan et al. 2021 . Among these MLPs , the NEP approach is known for its exceptional computational efficiency, enabled by its optimized formulation and efficient implementation in the graphics processing units molecular dynamics (GPUMD) package Fan et al. 2022 ; Fan et al. 2017 . Due to its high computational efficiency, the NEP approach has been extensively used in MD simulations, addressing complex challenges that traditional force fields typically can not tackle. Notably, the NEP method has been successfully applied in a wide range of material systems and research areas Dong et al. 2024 ; Ying et al. 2025 , such as predicting the structural properties of liquid and amorphous materials Wang et al. 2024 ; Xu et al. 2024 , unraveling chemical order in complex alloy systems Chen et al. 2024 , exploring phase transitions Fransson et al. 2023 ; Wiktor et al. 2023 , capturing surface reconstructions and material growth Qian et al. 2024 , simulating primary radiation damage Liu et al. 2023 , investigating fracture in two-dimensional materials Ying et al. 2023 ; Yu et al. 2024 , nanoscale tribology Ying et al. 2025 , as well as elucidating the mechanical behavior of compositionally complex alloys under various loading conditions Song et al. 2024 . Currently, a variety of tools have been developed for MLPs , offering functionalities such as training, evaluation, inference, fine-tuning, and deployment. For example, metatrain is a command-line interface (CLI) for training and evaluating atomistic models with diverse architectures, including the GAP metatensor 2025 . apax is a flexible and efficient open-source software package for both training and inference of MLPs , such as the Gaussian Moment Neural Network model Schäfer et al. 2025 . equitrain Benner 2025 provides a unified framework for training and fine-tuning MLPs . Similarly, mlip is a Python library for training and deploying MLPs Brunken et al. 2025 . In addition to these general-purpose toolkits, several automated frameworks have recently emerged to facilitate the generation of MLPs . For instance, autoplex provides a fully automated solution for creating high-quality MLPs , with support for the GAP Liu et al. 2024 . DP-GEN is dedicated to the automated generation and training of deep potential models Zhang et al. 2020 . AtomProNet Galib et al. 2025 is an open-source Python package that automates the acquisition of atomic structures, the preparation and submission of ab initio calculations, and the efficient collection of batch-processed data for streamlined neural network (NN) training, supporting models such as MACE Batatia et al. 2022 . However, existing tools for training MLPs rarely support architectures such as the NEP , and automated frameworks for fully training NEP models remain scarce. Notably, a key aspect in the development of MLPs is the generation of high-quality datasets for model training. Although recent toolkits such as calorine Lindgren et al. 2024 provide approaches for generating initial training sets, including strained, deformed, and rattled structures that can be used for the construction and training of NEP models, the preparation of NEP training datasets still heavily depends on user expertise. At present, few tools are available that can both systematically construct high-quality training structures and enable automated NEP model training. Moreover, graphical user interface (GUI) solutions for efficient screening and visualization of high-quality training datasets are also lacking. Addressing these limitations is the focus of our work, aiming to bridge this gap and further promote the broader application of NEP models. Therefore, in this work, we have developed NepTrain and NepTrainKit. NepTrain integrates the complete workflow, including dataset creation, training, and automated active learning, with built-in logic functions to constrain force ranges and filter out non-physical structures based on bond lengths. Furthermore, it allows seamless integration with Python and Bash scripts, enabling user-defined training workflows while automating processes through the use of NEP , GPUMD , Vienna ab initio simulation package (VASP) Kresse and Hafner 1994 ; Kresse and Furthmüller 1996 , and other tools. In parallel, NepTrainKit provides a user-friendly and powerful interactive GUI that enables researchers to visually inspect and edit training datasets. Its design aims to streamline the dataset construction process, enhance efficiency, and ensure the accuracy and reliability of potentials. The intuitive operation of NepTrainKit has significantly advanced the development of NEP and accelerates the study of material properties. This paper is organized as follows. In Section 2 , we introduce the NEP method and its

Read the complete text