Photoelectrochemical Properties, Machine Learning, and Symbolic Regression for Molecularly Engineered Halide Perovskite Materials in Water

The machine learning techniques are capable of predicting virtual material design space and optimizing material fabrication parameters. In this article, we construct machine learning models to describe the photoelectrochemical properties of molecularly engineered halide perovskite materials based on CH3NH3PbI3 in an aqueous solution and predict a complex multidimensional design space for the halide perovskite materials. The machine learning models are trained and tested based on an experimental photocurrent data set consisting of 360 data points with varying experimental conditions and dye structures. Machine learning algorithms including support vector machine (SVM), random forest, k-nearest neighbors, Rpart, Xgboost, and Kriging algorithms are compared, with the Kriging algorithm achieving the best accuracies (r = 0.99 and R-2 = 0.98) and SVM achieving the second best. A total of 50,905 data points representing the complex multidimensional design space are predicted via the machine-learned models to benefit the future perovskite studies. In addition, the symbolic regression based on the genetic algorithms effectively and automatically designs hybrid descriptors that outperform the individual descriptors. This article highlights the machine learning and symbolic regression methods for designing stable and highperformance halide perovskite materials and serves as a platform for further experimental optimization of halide perovskite materials.