Machine learning potential for the Cu-W system

Combining the excellent thermal and electrical properties of Cu with the high abrasion resistance and thermal stability of W, Cu-W nanoparticle-reinforced metal matrix composites and nano-multilayers (NMLs) are finding applications as brazing fillers and shielding material for plasma and radiation. Due to the large lattice mismatch between fcc Cu and bcc W, these systems have complex interfaces that are beyond the scales suitable for ab initio methods, thus motivating the development of chemically accurate interatomic potentials. Here, a neural network potential (NNP) for Cu-W is developed within the Behler-Parrinello framework using a curated training dataset that captures metallurgically-relevant local atomic environments. The Cu-W NNP accurately predicts (i) the metallurgical properties (elasticity, stacking faults, dislocations, thermodynamic behavior) in elemental Cu and W, (ii) energies and structures of Cu-W intermetallics and solid solutions, and (iii) a range of fcc Cu/bcc W interfaces, and exhibits physically-reasonable behavior for solid W/liquid Cu systems. As will be demonstrated in forthcoming work, this near-ab initio-accurate NNP can be applied to understand complex phenomena involving interface-driven processes and properties in Cu-W composites.

结合铜（Cu）优异的热学与电学性能以及钨（W）的高耐磨性与热稳定性，铜-钨（Cu-W）纳米颗粒增强金属基复合材料与纳米多层膜（nano-multilayers, NMLs）已被应用为钎料及等离子体与辐射屏蔽材料。由于面心立方（face-centered cubic, fcc）铜与体心立方（body-centered cubic, bcc）钨之间存在较大晶格失配，该体系的复杂界面尺度超出了从头算方法（ab initio methods）的适用范围，因此推动了化学精准原子间势（interatomic potentials）的开发。本研究基于贝勒-帕里内洛框架（Behler-Parrinello framework），利用覆盖冶金相关局域原子环境的精选训练数据集，构建了适用于Cu-W体系的神经网络势（neural network potential, NNP）。该Cu-W神经网络势可精准预测以下内容：(i) 单质铜与钨的冶金性能（弹性、层错、位错及热力学行为）；(ii) 铜-钨金属间化合物与固溶体的能量与结构；(iii) 多种面心立方铜/体心立方钨界面，并在固态钨/液态铜体系中表现出符合物理规律的行为。正如后续研究将展示的那样，这种接近从头算精度的神经网络势可用于解析铜-钨复合材料中涉及界面驱动过程与性能的复杂现象。