Machine learning prediction of DFT-derived electrical conductivity in perovskite and double perovskite materials using physically motivated descriptors.
TL;DR
Evaluated classical machine-learning models trained on different descriptor sets including DFT band gap, non-orbital compositional descriptors, orbital-related descriptors, SOAP structural fingerprints, and a reduced mixed descriptor set to predict DFT-derived transport conductivity demonstrate that accurate prediction can be achieved using a small number of physically motivated variables.
Abstract
Predicting electrical conductivity in perovskite and double-perovskite materials remains challenging as this property depends on various electronic, chemical, and structural factors. In this work, we evaluate classical machine-learning models trained on different descriptor sets including DFT band gap, non-orbital compositional descriptors, orbital-related descriptors, SOAP structural fingerprints, and a reduced mixed descriptor set to predict DFT-derived transport conductivity. The band gap provides a strong baseline but is insufficient to fully predict the target. The best overall performance is obtained using non-orbital compositional descriptors with Random Forest regression, while orbital-related descriptors achieve nearly comparable accuracy, confirming the importance of valence-electron characteristics. A compact mixed descriptor set preserves nearly the full predictive power of the larger descriptor spaces, showing that accurate prediction can be achieved using a small number of physically motivated variables. In contrast, SISSO showed much lower accuracy, suggesting that sparse symbolic expressions are insufficient to capture the nonlinear relationships underlying our target values. These results demonstrate that physically informed classical machine-learning models can provide an effective surrogate framework for reproducing DFT/BoltzTraP-derived conductivity trends in perovskite and double-perovskite materials.