A domain knowledge enhanced machine learning method to predict the properties of halide double perovskite A2B+B3+X6

Abstract

The heat of formation and band gap are important properties of halide double perovskites, which can dictate the range of application. However, traditional laborious experiments and density functional theory (DFT) calculations are time-consuming, and machine learning methods perform poorly in acquiring hidden knowledge from material samples, which make it difficult to explore promising perovskites with high stability efficiently. Here, we present a new machine learning framework, which could embed material domain knowledge into a machine learning method, and then it is used to predict the heat of formation and band gap. Compared with traditional machine learning methods, we found that the prediction accuracy of the heat of formation and band gap in 540 perovskite materials A₂B⁺B³⁺X₆ is improved by 5.65% and 15.52% (namely 93.58% and 62.65%) respectively. It is shown that material domain knowledge contains a wealth of logical rules and expert experience, which can remedy most of the problems in data-driven materials research, such as data sparse and complex relationships between features.