Research progress and perspective of machine learning in material design

With rapid development of economy and society, the excessive demand for resources has caused the imbalance of ecological environment. It is therefore urgent to develop new functional materials, especially energy conversion materials, to solve the scientific and engineering problems in the field of resources and environment. However, the research and development of materials was traditionally based on inefficient trial-and-error experiments. Although state-of-the-art approach such as density functional theory (DFT) is able to simulate materials properties, the calculations of high temperature, high pressure and strong magnetic field environment, as well as the selection of strong and weak correlation system between electrons and interaction potential between atoms are still unsatisfactory. Huge amount of data produced by experiments and simulations provides databases for machine learning. Combining the theory of probability and statistics algorithm, machine learning has recently made much progress in the new material discovery and design, the prediction of material performance and application and other purposes ranged from the macroscopic to the microscopic scale, such as the statistics classification of perovskite materials, the stability prediction of perovskite materials based on high-throughput computing and intermetallic compound electrocatalysts design and selection, etc. Meanwhile, developing a physically interpretable descriptor that captures the trend of materials properties is a critical goal of data-driven science. Machine learning has been applied in the field of materials science and engineering, exhibiting a different perspective from traditional approaches. In this paper, recent progress of materials design in photovoltaics, electrocatalytic and performance evaluation of energy storage batteries are reviewed. In those efforts, machine learning aims to discover the relationships among compositional and structural features and functionality in complex systems of materials. Machine learning is a data-driven approach that relies heavily on data. Compared with image recognition and other fields that usually have millions of data, material science research often leads to over-fitting of machine learning models when the amount of training data is limited, which greatly reduces the generalization ability of machine learning methods. In order to increase the amount of material data, researchers can obtain theoretical data through high-throughput computing on the one hand, and develop methods for intelligent reading of literatures to access and obtain a large number of relevant experimental and theoretical data from publications on the other hand. Another promising method to solve the problem of finite data sets is meta-learning, that is, learning knowledge within or across problems. The development of new technologies such as neural Turing machine and imitation learning makes this process possible. Recently, it has been reported that Bayesian optimization can reach the experience level of human judgment on things through one-shot learning under the condition of limited data, which may have a huge promotion effect on materials science with scarce data with slow and expensive acquisition speed. Although machine learning methods have greatly improved their predictive accuracy in material discovery, design, performance and application, they have not expanded well in terms of transferability. Active learning methods also provide consistent and automated improvements in accuracy and transferability, making a significant contribution to the success of the universal model. In addition, one of the promising point of material science using machine learning method is to develop new descriptor owning physical interpretability, makes the black box model of statistical machine learning be explainable. In a word, the development of computer intelligence algorithms would promote the innovation of new materials discovery.