机器学习在矿物结构搜索及性质预测方面的应用

Applications of machine learning in the mineral structure search and mineral property prediction

研究矿物在高温高压下的结构及其物理、化学性质是了解地球内部物质组成及动力学演化的基础。实验观测和理论计算是目前最主要的两种方法,前者受技术的局限,高温高压下的数据点比较匮乏;后者容易实现高温高压,但计算精度和效率往往难以两全。基于上述问题,机器学习作为一种很有前景的工具,与第一性原理计算相结合,能够精确、高效地预测矿物结构及各种性质。本文首先介绍了构建机器学习势的一般流程,对目前应用广泛的深度势能方法进行详细阐述。然后探讨了机器学习方法在计算矿物物理领域中的应用,如矿物新相结构搜索、热力学性质和输运性质预测、元素配分及同位素分馏等,并分析了机器学习方法相比于传统方法的优势。最后对机器学习方法目前存在的不足进行了简单总结并对其应用前景提出展望。

Investigating the structures as well as physical and chemical properties of minerals under high temperature and high pressure is the key basis for understanding the composition and dynamic evolution of the Earth’s interior matters. The experimental observation and theoretical calculation are two main currently applied methods. Due to the limitation of technology on the experiments, experimental data under high temperature and high pressure conditions are scarcely obtained. Although it is easy for the theoretical calculations to achieve relevant data under high P-T conditions, it is difficult to obtain both of the high accuracy and high efficiency for the calculations. Based on currently existed problems, the machine learning as a promising tool has been combined with the first-principles calculations for accurately and efficiently predicting mineral structures and properties. In this paper, we have firstly introduced the general process of constructing a machine learning potential and have elaborated in details the Deep Potential method which is widely used at present. Then we have discussed applications of the machine learning method in the field of computational mineral physics such as the searching for structures of new mineral phases, the prediction of thermodynamic and transport properties, and the element partition and isotope fractionation, and have analyzed advantages of the machine learning relative to traditional methods. Finally, we have briefly summarized the currently existed limitations of the machine learning method and have provided an outlook for its future applications.