Bandgap model using symbolic regression for environmentally compatible lead-free inorganic double perovskites

Data-driven models have become an essential practice of scientific research in the perovskite field, along with theory and experiments. Material informatics has emerged as a viable alternative method of exploring and formulating novel perovskite compounds using a descriptor-based approach. Herein, we develop a method that includes feature augmentation with symbolic regression to rapidly estimate and screen non-toxic lead-free inorganic double perovskites (A 2 BB'X 6 ) using machine learning. Predictive models were created by identifying a physico-chemical relevant descriptor from an extensive pool of augmented features. Using primary atomic and molecular features, a high dimensional space of descriptors $(\text{containing}\approx 3\times 10^{5}$ features) was reconstructed using mathematical operators. By increasing the complexity from 1-D to 5-D descriptor, the correlation coefficient was increased from 81.6% to 92.4%. These accurate and interpretable models can then be employed for screening lead-free perovskites with appropriate bandgaps and stability.