Regression Models for Lattice Thermal Conductivity Prediction

RAHUL GHOSH, NABENDU MONDAL

Keywords: Regression Model, Lattice Thermal Conductivity

Abstract:

Predicting lattice thermal conductivity is vital for designing efficient thermoelectric materials, but conventional first-principles methods such as DFPT are computationally demanding. In this work, we apply machine learning (ML) techniques to predict lattice thermal conductivity using a curated AFLOW dataset of ~5,500 compounds at 300 K. Physics-informed features, including compositional, stoichiometric, and chemical descriptors, are engineered and refined through feature selection methods like RFE and PCA. Multiple regression models, including Linear Regression, KNN, Decision Tree, Random Forest, and Boosting, are trained and validated. Our study demonstrates that ML can bypass costly calculations, enabling accurate predictions and accelerating the discovery of thermoelectric materials with low lattice thermal conductivity.

https://youtu.be/xDlYNLb1R1Y

Submission Number: 3