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

Zeyu Wang, Shuya Yamazaki, Martin Hoffmann Petersen, Masato Ohnishi, Tomiya Yamamoto, Wei Nong, Jianghai Wang, Ruiming Zhu, Masatoshi Hanai, Michimasa Morita, Toyotaro Suzumura, Zekun Ren*, Junichiro Shiomi*, Kedar Hippalgaonkar* · 2026

Read Fast and Accurate Prediction of Lattice Thermal Conductivity via Machine Learning Surrogates by Zeyu Wang, Shuya Yamazaki, Martin Hoffmann Petersen, Masato Ohnishi, Tomiya Yamamoto, Wei Nong, Jianghai Wang, Ruiming Zhu, Masatoshi Hanai, Michimasa Morita, Toyotaro Suzumura, Zekun Ren*, Junichiro Shiomi*, Kedar Hippalgaonkar* (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-11610.txt.

Introduction

Wang and thirteen others benchmark 15 stand-in models for lattice thermal conductivity. New models have opened vast matter spaces, but high-quality thermal conductivity work still needs exact anharmonic treatment. Using the Phonix base of 6,966 entries from theory work, they group stand-ins into physics-aware features, end-to-end deep networks, and trained MLIP codes. MLIP-based models excel at filling gaps, while deep networks prove most robust on new data. Stand-ins cut compute costs many times over for high-throughput heat-power screening.

Catalog record

CREATORZeyu Wang, Shuya Yamazaki, Martin Hoffmann Petersen, Masato Ohnishi, Tomiya Yamamoto, Wei Nong, Jianghai Wang, Ruiming Zhu, Masatoshi Hanai, Michimasa Morita, Toyotaro Suzumura, Zekun Ren*, Junichiro Shiomi*, Kedar Hippalgaonkar*
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48479
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.11610

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fast and Accurate Prediction of Lattice Thermal Conductivity via Machine Learning Surrogates Zeyu Wang, Shuya Yamazaki, Martin Hoffmann Petersen, Masato Ohnishi, Tomiya Yamamoto, Wei Nong, Jianghai Wang, Ruiming Zhu, Masatoshi Hanai, Michimasa Morita, Toyotaro Suzumura, Zekun Ren*, Junichiro Shiomi*, Kedar Hippalgaonkar* 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 κ l ​ a ​ t \kappa_{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 κ l ​ a ​ t \kappa_{lat} using the Phonix database [ 28 ] , 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 κ l ​ a ​ t \kappa_{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- κ l ​ a ​ t \kappa_{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. 1 Introduction In recent years, Machine Learning (ML) has emerged as a powerful tool in computational materials science [ 33 , 3 , 26 , 34 , 4 , 7 , 29 ] . In particular, Machine Learning Interatomic Potentials (MLIPs) have enabled near–first-principles accuracy at significantly reduced computational cost [ 47 , 1 , 5 , 40 ] , while advances in generative models have enabled the exploration and design of materials in large and complex chemical spaces [ 49 , 13 , 15 , 44 ] . The introduction of generative models has initiated a new paradigm: rather than relying on chemically intuitive, combinatorial searches for promising functional materials, we can now train generative models on descriptive databases and sample candidate materials directly from the learned distributions [ 22 ] . Although these models can be conditioned on specific chemical spaces or targeted properties, ensuring that the generated structures faithfully satisfy the desired properties typically requires extensive post hoc evaluation. A key example of functional materials of interest is thermoelectrics, which can convert waste heat directly into electricity or operate as solid-state Peltier coolers. Despite their technological potential, thermoelectric materials and devices still have limited practical adoption due to challenges related to efficiency, material cost, and large-scale manufacturability [ 35 , 42 ] . The efficiency of a thermoelectric material is quantified by the dimensionless figure of merit ZT=S^{2}\sigma T/(\kappa_{lat}+\kappa_{el}) , where S is the Seebeck coefficient, \sigma electric conductivity, \kappa_{lat} the lattice thermal conductivity, \kappa_{el} the electric thermal conductivity, and T the absolute temperature [ 43 , 37 ] . Achieving high thermoelectric performance therefore requires maximizing the power factor, S^{2}\sigma while simultaneously minimizing the total thermal conductivity. Since most high-performance thermoelectric materials are heavily doped semiconductors [ 24 ] , where phonon-mediated heat transport dominates over the electronic contribution, the lattice thermal conductivity \kappa_{lat} typically represents the major component of the total thermal conductivity. As a result, accurately predicting \kappa_{lat} is crucial for screening and optimizing thermoelectric materials [ 37 , 10 ] . Computing lattice thermal conductivity from first-principles typically requires either constructing large supercells to extract harmonic and anharmonic interatomic force constants (IFCs), which capture both phonon dispersions and three-phonon or four-phonon scattering processes, and subsequently solving the phonon Boltzmann transport equation (BTE) or performing long ab-initio molecular dynamics (AIMD) simulations to capture heat transport over extended timescales [ 14 , 9 ] . Both approaches are computationally demanding, and even when first-principles calculations are substituted with MLIPs, the computational cost remains prohibitive, rendering large-scale screening of the extensive structural space produced by generative models infeasible. To address this challenge, we evaluate a diverse set of surrogate models for predicting lattice thermal conductivity directly from given candidate structures. These include end-to-end deep neural networks that take atomic structures as input and learn task-specific representations through their architectures [ 41 , 46 ] , feature-engineered models that rely on hand-crafted physicochemical descriptors with strong inductive bias [ 48 ] , and models that leverage pre-trained universal MLIPs embeddings, which are subsequently used as input features to train downstream feedforward neural networks. In this study, we create a benchmark by developing 15 surrogate models using the Phonix dataset [ 27 ] , which contains 6,966 entries with anharmonic phonon properties derived from first-principles calculations. Owing to its broad chemical and structural coverage, this dataset enables surrogate models to learn relationships between general structures and thermal transport properties, allowing rapid identification of low lattice thermal conductivity candidates and robust generalization beyond materials traditionally studied for thermoelectric applications. To systematically assess model performance and generalization, we train and evaluate each model under three dataset splits designed to probe qualitatively different aspects of predictive capability. First, a random split serves as the standard baseline, where training and test materials are drawn from the same overall distribution, assessing general interpolation accuracy. Second, a space-group disjoint split ensures that all crystal symmetries (space groups) present in the test set are entirely absent from the training set; this reflects a realistic scenario in which models are deployed on structurally novel materials and directly tests the ability to extrapolate across unseen structural motifs. Third, an out-of-distribution (OOD) split based on \kappa_{lat} trains models exclusively on high- \kappa_{lat} materials ( \kappa_{lat} > 1 \mathrm{Wm^{-1}K^{-1}} ) and evaluates them on low- \kappa_{lat} materials ( \kappa_{lat} \leq 1 \mathrm{Wm^{-1}K^{-1}} ). This threshold is motivated by the thermal conductivity requirements of state-of-the-art thermoelectric materials, for instance, Bi 2 Te 3 achieves zT > 1 at room temperature in part due to its intrinsically low- \kappa_{lat} of \sim 1 \mathrm{Wm^{-1}K^{-1}} . This design simulates the realistic challenge of generalizing from the more abundant, high-conductivity materials in existing datasets to the rare, low-conductivity candidates most desirable for thermoelectric applications. Together, these three splits enable us to probe general predictive accuracy, structural extrapolation, and property-based extrapolation, respectively.The best-performing models are identified as the most promising candidates for integrating with generative design workflows for accelerated thermoelectric materials discovery. Figure 1 : Material distribution in Phonix. (a) Distribution and percentage of crystal systems. (b) Distribution of \kappa_{lat} at 300 K, derived from the Phonix database [ 27 ] . The subgraph represents the percentage of materials with different \kappa_{lat} . Low \kappa_{lat} means that thermal conductivity values are lower than 1 \mathrm{Wm^{-1}K^{-1}} . (c) Umap for comparing with all Materials Project structures [ 12 ] . 2 Methods 2.1 Lattice

Read the complete text