基于PyTorch的晶体结构势预测

Prediction for Potentials of Crystal Structures Based on PyTorch

近年来,材料数据进入了爆发式增长阶段,而实验和模拟方法来处理和分析材料数据工程量巨大并且十分耗时,如何用新方法从大量的数据中发现知识是未来材料研发的主要探索方向。新世纪以来,出现了将机器学习方法应用于材料科学领域研究的新课题。高效准确地得到结构能量是许多材料研究工作的基础,我们主要关注将机器学习方法与结构能量计算相结合这一方向。本文调研了目前已有的基于机器学习的结构势能拟合开源软件,分析它们的特点与优劣,并针对实际需求通过深度学习框架PyTorch对软件进行优化,增强机器学习模型的拟合能力。PyTorch可以利用GPU资源对数据进行快速学习,灵活地利用多种优化算法对不同体系的结构搭建合适的模型,并将模型保存以供后续研究。我们在人工智能计算及数据应用服务平台上对软件进行测试与优化,解决了一些在测试中发现的问题,对不同的体系结构的结构势能拟合效果进行了分析,提升软件对不同体系的结构势能拟合能力。

In recent years, data in the field of materials has entered an explosive growth stage, and analyzing the data with experimental and simulation methods are time consuming. How to use new methods to discover knowledge from a large amount of data is the main exploration direction for future research anddevelopment. Since the new century, new topics about machine learning methods in the field of materials science have emerged. Obtaining structural energy efficiently and accurately is the basis of many materials research. We mainly focus on the way to use machine learning methods in structural energy calculations. This paper investigates the existing ML-based open source software for structural potential energy fitting, analyzes their advantages and disadvantages, and optimizes the software through the deep learning framework PyTorch to enhance the fitting ability of the machine learning model. PyTorch can use GPU resources to learn from data quickly, use multiple optimization algorithms to build appropriate models flexibly for different systems with different atom types, and save the model for subsequent research. We tested and optimized our software on the artificial intelligence computing and data service platform, solved some problems found in the testing process, analyzed the results of potential energy fitting for different systems and improved the fitting ability of software based on them.