A physics-informed neural network for temperature-dependent lattice thermal conductivity in half-Heusler compounds

Lattice thermal conductivity ( κ l ) is an essential property for designing efficient thermoelectric materials, yet fully anharmonic lattice dynamics calculations are computationally demanding. In this work, we developed a physics-informed neural network to predict full temperature-dependent lattice thermal conductivity, κ l ( T ), by combining an empirical baseline and a correction term. The model was trained on first-principles κ l data of 65 half-Heusler compounds over the temperature 200–1000 K, using only elemental and structural descriptors. On the test compounds, the model achieves high accuracy ( R 2 = 0.94 ), demonstrating strong generalization across both composition and temperature. Notably, when evaluated at a single operating point (300 K), it maintains high predictive performance ( R 2 = 0.91 ), showing that it remains reliable for room temperature screening. Furthermore, we apply the model to screen a database of half-Heusler compounds and rank candidates across the full operating window rather than at one reference temperature, enabling temperature-aware selection of compositions for thermoelectric applications.