npj Computational Materials

Machine learning force fields for molecular liquids: Ethylene Carbonate/Ethyl Methyl Carbonate binary solvent

Ioan-Bogdan Magdău, Daniel J. Arismendi-Arrieta, Holly E. Smith, Clare P. Grey, Kersti Hermansson & Gábor Csányi

npj Computational Materials 9: 146 (2023)

doi.org/10.1038/s41524-023-01100-w

Published online: 17 August 2023

编辑概述

机器学习引领电池分子液体模拟之巅

高精度的从头算分子动力学方法在研究凝聚相分子机制中发挥着关键作用，然而，由于其昂贵的计算成本，难以涵盖许多关键性质。机器学习方法为模拟大长度尺度和长时间尺度提供了可能性，但在分子凝聚相中尺度分离和复杂的分子混合物中仍面临挑战。该研究通过在总相互作用上拟合机器学习模型，开发一种适用于分子液体混合物的通用、准确且经济的机器学习力场。

来自英国剑桥大学工程实验室的Ioan-Bogdan Magdău教授团队，采用了一种独特的机器学习方法，即在总相互作用上拟合模型，避免了传统分子模拟中尺度分离的问题。通过迭代训练和详细测试，针对EC:EMC二元溶剂，即锂离子电池中的液体电解质关键组分，开发了高精度机器学习势。他们通过总损失函数的拟合，实现了在分子内尺度上的良好准确度；通过采用了刚性分子体积扫描和分子内/分子间误差分离等技术，改进了模型在分子间尺度上的性能。对多组分分子液体的建模展示了对各种局部分子环境的均匀表示的必要性。值得注意的是，采用短程模型来描述中性溶剂液体，在一定程度上简化了模型。该研究为未来电池系统等领域的机器学习建模提供了经验，为全反应力场的发展提供了新的方法和思路。

Editorial Summary

Machine learning force fields：Molecular liquids in batteries
High-precision ab initio molecular dynamics methods play a crucial role in studying the molecular mechanisms of condensed-phase systems. However, due to their expensive computational costs, they struggle to cover many crucial properties. Machine learning methods offer the possibility of simulating large length and time scales, but face challenges in scale separation and complex molecular mixtures within molecular condensed phases. A team led by Prof. Ioan-Bogdan Magdău's from the Engineering Laboratory, the University of Cambridge, introduced a unique machine learning approach—fitting the model on the total interaction—to address the scale separation issue present in traditional molecular simulations. Through iterative training and detailed testing, they developed a high-precision machine learning potential for the EC:EMC binary solvent, a critical component in the liquid electrolyte of lithium-ion batteries. Achieving good accuracy at the molecular scale by fitting the total loss function, they improved the model's performance at the intermolecular scale by employing techniques such as rigid-molecule volume scans and intra-/inter-error separation. Modeling multi-component molecular liquids highlighted the necessity of a uniform representation of various local molecular environments. Notably, adopting a short-range model to describe neutral solvent liquids simplified the model to some extent. This research provides valuable experience for machine learning modeling in future fields like battery systems and offers new insights and approaches for the development of comprehensive reactive force fields.