Machine Learning-Accelerated Prediction of Thermal Conductivity in Layered Perovskite Structures

Abstract

Layered perovskite materials exhibit thermal transport properties relevant to thermoelectric and optoelectronic applications, yet first-principles calculation of thermal conductivity across large compositional spaces remains computationally expensive, limiting systematic materials screening. This study develops a machine learning framework trained on a curated dataset of density functional theory calculations to predict lattice thermal conductivity across a range of layered perovskite compositions with varying layer thickness and cation substitution. The model incorporates structural descriptors capturing octahedral distortion and interlayer spacing alongside compositional features, allowing predictions to generalize across compositional variations not present in the training set. Validation against held-out first-principles calculations shows prediction accuracy sufficient to reliably rank candidate compositions by relative thermal conductivity, substantially reducing the computational cost of preliminary screening compared to exhaustive first-principles evaluation. We apply the trained model to screen a broader candidate space, identifying several compositions predicted to exhibit notably low thermal conductivity consistent with favorable thermoelectric performance, which we recommend for targeted experimental follow-up. The approach demonstrates how machine learning surrogate models can meaningfully accelerate early-stage materials discovery workflows without eliminating the role of first-principles validation for final candidate selection.