LixTMO2 (TM=Ni, Co, Mn) forms an important family of cathode materials for Li-ion batteries, whose performance is strongly governed by Li composition-dependent crystal structure and phase stability. Here, we use LixCoO2 (LCO) as a model system to benchmark a machine learning-enabled framework for bridging scales in materials physics. We focus on two scales: (a) assemblies of thousands of atoms described by density functional theory-informed statistical mechanics, and (b) continuum phase field models to study the dynamics of order-disorder transitions in LCO. Central to the scale bridging is the rigorous, quantitatively accurate, representation of the free energy density and chemical potentials of this material system by coarsegraining formation energies for specific atomic configurations. We develop active learning workflows to train recently developed integrable deep neural networks for such high-dimensional free energy density and chemical potential functions. The resulting, first principles-informed, machine learning-enabled, phase-field computations allow us to study LCO cathodes' phase evolution in terms of temperature, morphology, charge cycling and particle size.

翻译：LixTMO₂（TM=Ni, Co, Mn）是锂离子电池中一类重要的正极材料，其性能强烈受锂组分依赖的晶体结构和相稳定性控制。本文以LixCoO₂（LCO）为模型系统，评估了一种基于机器学习的框架在材料物理学中连接尺度的能力。我们聚焦于两个尺度：（a）由密度泛函理论统计力学描述的包含数千个原子的集合体，以及（b）用于研究LCO中有序-无序转变动力学的连续相场模型。尺度连接的核心在于，通过粗粒化特定原子构型下的形成能，对这一材料体系的自由能密度和化学势进行严格且量化精确的表征。我们开发了主动学习流程，以训练近期提出的可积分深度神经网络，用于高维自由能密度和化学势函数。由此获得的第一性原理指导、机器学习赋能的相场计算，使我们能够研究LCO正极在不同温度、形貌、充放电循环及颗粒尺寸下的相演化过程。