Fast and Accurate Prediction of Lattice Thermal Conductivity via Machine Learning Surrogates

TL;DR

This paper benchmarks 15 machine learning surrogate models for predicting lattice thermal conductivity, highlighting their strengths and limitations in interpolation and extrapolation, and emphasizing their potential for high-throughput materials screening.

Contribution

It provides a comprehensive comparison of diverse ML surrogate models for ppatl prediction, including new insights into their generalization capabilities and performance in out-of-distribution scenarios.

Findings

MLIP-embedded models excel in interpolation within known data regions.

Deep neural networks like ALiEGNN perform best in out-of-distribution extrapolation.

Surrogate models significantly reduce computational costs compared to first-principles calculations.

Abstract

The appearance of generative models has opened vast chemical spaces in the design of functional materials. Although machine learning interatomic potentials (MLIPs) have substantially accelerated phonon calculations, high-fidelity prediction of lattice thermal conductivity \k{appa}lat still requires accurate treatment of anharmonic interactions, which remains a key challenge for existing potentials across novel chemical spaces. To address this challenge, we present a comprehensive benchmark of 15 surrogate models for predicting \k{appa}lat using the Phonix database, which contains 6,966 entries with anharmonic phonon properties derived from first-principles calculations. Firstly, We categorize these surrogate models into three distinct groups: Physical-informed feature descriptors combined with ML models, end-to-end deep neural networks, and pre-trained MLIP-embeddings combined with ML models. By evaluating model performance across random, space-group disjoint (testing generalization to unseen crystal symmetries), and Out-Of-Distribution splits (OOD dataset that testing extrapolation to property regimes beyond the training range) based on \k{appa}lat, we probe both interpolation and exploration capabilities. Our results reveal that MLIP-embedded models excel in interpolation within well-sampled regions, deep neural network models especially ALiEGNN demonstrate superior robustness in OOD regimes critical for discovering novel low-\k{appa}lat. Additionally, we find a systematic degradation in performance when the structural representation is reduced. Although surrogate models exhibit lower accuracy than direct simulations using first-principles calculation, they reduce computational costs by orders of magnitude, enabling efficient high-throughput screening of thermoelectric materials with minimal loss in generative design workflows.