Evaluation of Machine Learning Models in the Search of Interpretable Structure Maps of Perovskites and Spinels through Modified SISSO Routine

The systematic classification of functional inorganic materials, such as perovskites and spinels, is a prerequisite for accelerating the discovery of advanced electronic and optoelectronic devices. This type of classification was performed for several binary materials in the form of structure maps, which has interpretable geometrical or electronic properties of the elements as indices. Lately, machine learning has been used increasingly to extend this classification, most of them comes at the cost of interpretability due to the black-box nature of the models used. Symbolic classification offers a path to interpretable models by creating equations composed of starting elemental properties as features, however, the selection of best equations needs to be robust to handle common issues associated with the experimental data. This study examines the use of several machine learning models for ranking the equations produced by Sure Independence Screening and Sparsifying Operator (SISSO). The scoring metrics show that the K-nearest neighbor (KNN) approach as optimal for distinguishing perovskites with an accuracy of 0.92 (20% improvement over the Goldschmidt tolerance factor) and spines with an accuracy of 0.75. Moreover, compared to the other models tested, the KNN-modified SISSO routine was selected as the best model in 8 out of 10 cases among the top thousand pairs.