Development of an international, large-scale database of thermal conductivity calculated by first principles

Advancing the discovery of thermal functional materials using high-precision data and AI

Key Points

• An international research team developed “auto-kappa,” software that automates first-principles calculations, and constructed “Phonix,” a large-scale database of thermal conductivity and phonon properties for more than 6,800 inorganic crystalline materials.
• Using this database, the team developed a graph neural network-based machine learning model and discovered that a “scaling law” holds, whereby prediction accuracy improves exponentially as the amount of training data increases.
• The database and AI model were demonstrated to enable the efficient discovery of novel materials with extremely high or extremely low thermal conductivity. The application of such AI for Science approaches is expected to further accelerate materials development, including thermal management materials for next-generation electronic devices and thermoelectric materials.

Materials discovery using the international, large-scale thermal conductivity database

Abstract

An international collaborative research group consisting of researchers from around the world—including Professor Junichiro Shiomi of the Graduate School of Engineering, The University of Tokyo; Project Associate Professor Masato Ohnishi of the Institute of Statistical Mathematics (also Visiting Researcher at the Graduate School of Engineering, The University of Tokyo); Professor Ryo Yoshida of the same institute; Group Leader Terumasa Tadano of the National Institute for Materials Science; Professor Toyotaro Suzumura of the Graduate School of Information Science and Technology, The University of Tokyo (also affiliated with the Information Technology Center, The University of Tokyo); Project Assistant Professor Masatoshi Hanai of the Information Technology Center, The University of Tokyo; Professor Tengfei Luo of the University of Notre Dame; Associate Professor Kedar Hippalgaonkar of Nanyang Technological University; Professor Alan McGaughey of Carnegie Mellon University; Senior Researcher Lucas Lindsay of Oak Ridge National Laboratory; Professor Xiulin Ruan of Purdue University; Professor Ming Hu of the University of South Carolina; and Professor Zhiting Tian of Cornell University—developed an automated computational system for analyzing anharmonic interactions of phonons (Note 1), which determine thermal transport in crystalline materials, based on first-principles calculations (Note 2). They also constructed a large-scale database, “Phonix,” containing lattice thermal conductivity and phonon properties for more than 6,800 inorganic materials (Figure 1).

Figure 1: Materials discovery using the international, large-scale thermal conductivity database

Phonons are quasiparticles that carry heat in solids, and anharmonic phonon properties (Note 3), such as thermal conductivity and phonon lifetimes, have a significant impact on the performance of functional materials, including heat dissipation materials and thermoelectric materials. In this study, we developed “auto-kappa” software that automates first-principles calculations, and computed phonon lifetimes and lattice thermal conductivity (Note 4) for more than 6,800 materials to construct a comprehensive database. The database also includes harmonic phonon properties (Note 5), such as phonon dispersion relations, for more than 12,000 materials.

Furthermore, using this database, we developed a machine learning (AI) prediction model based on graph neural networks (Note 6), and identified novel candidate materials exhibiting high and low thermal conductivity through large-scale screening of approximately 380,000 crystal structures. We also demonstrated that a “scaling law” (Note 7) holds, whereby prediction accuracy improves exponentially as the amount of training data increases. This important result suggests that further expansion of the dataset could significantly enhance the accuracy of material property predictions.

Another key feature of this work is that it was conducted as an international collaborative effort involving researchers from around the world. This is expected to promote the continued development, expansion, and utilization of the database as a global materials resource.

These results are expected to deepen our understanding of thermal transport and significantly advance data-driven discovery of thermal functional materials using AI.

Background, Achievements, and Prospects

In recent years, Materials Informatics (Note 8), which integrates data science and materials science, has rapidly advanced, accelerating research and development in areas such as battery materials, catalysts, and magnetic materials. A key foundation for this progress is large-scale materials databases. Open databases such as the Materials Project (Note 9) have accumulated first-principles calculation data for hundreds of thousands of materials. However, most of these datasets are limited to relatively easy-to-compute properties, such as crystal structures and electronic structures, and databases that systematically include anharmonic phonon properties—critical for determining thermal transport—have been largely unavailable. Anharmonic phonon properties govern lattice thermal conductivity and phonon lifetimes, playing essential roles in thermal management of electronic devices and the performance of thermoelectric materials. However, their computation requires substantial computational resources, posing a major barrier to large-scale database construction.

In this study, we developed “auto-kappa,” software that automates phonon analysis based on first-principles calculations. This software automates complex computational workflows, including crystal structure optimization, phonon dispersion calculations, and three-phonon scattering calculations, thereby enabling large-scale computations. Using this tool, we calculated phonon properties for more than 6,800 inorganic materials and constructed the database “Phonix,” which includes phonon dispersion relations, lifetimes, scattering rates, mean free paths, and lattice thermal conductivity.

Furthermore, using this database, we developed a machine learning model based on graph neural networks and demonstrated that lattice thermal conductivity can be predicted with high accuracy from crystal structures. By applying this model to the exploration of approximately 380,000 candidate crystal structures, we identified novel materials exhibiting extremely high or extremely low thermal conductivity.

A distinctive feature of this work is that it was conducted as an international collaborative effort driven by academic motivations, involving researchers from around the world. While various databases for materials and functionalities are being developed globally, it is often difficult to achieve stepwise global integration after establishing local databases. In contrast, initiating this effort as an open international collaboration from the outset provides significant advantages in terms of accelerating data generation and expanding the user base.

Looking ahead, we plan to extend the database by incorporating more complex interactions, such as four-phonon scattering and electron–phonon interactions, as well as by accelerating computations through machine learning. The data and machine learning models developed in this study are expected to serve as a foundation for AI for Science (Note 10), accelerating materials design across a wide range of applications, including heat dissipation materials for electronic devices and thermoelectric conversion materials.

〇 Related Resources:

auto-kappa software: https://github.com/phonix-db/auto-kappa

Phonix database: https://phonix-db.org/

Researchers

The University of Tokyo

  Graduate School of Engineering

    Junichiro Shiomi, Professor

Also: RIKEN Center for Advanced Intelligence Project, Visiting Researcher

Institute of Statistical Mathematics, Research Center for Advanced Data Science, Visiting Professor

    Zeyu Wang, PhD Student

    Michimasa Morita, PhD Student (at the time of the research)

Current: The University of Tokyo, Assistant Professor

  Graduate School of Information Science and Technology

    Toyotaro Suzumura, Professor

Also: Information Technology Center, The University of Tokyo, Professor

  Information Technology Center

    Masatoshi Hanai, Project Assistant Professor

National Institute for Materials Science

  Research Center for Magnetic and Spintronic Materials

    Terumasa Tadano, Group Leader

Institute of Statistical Mathematics

  Research Center for Advanced Data Science

    Masato Ohnishi, Project Associate Professor

Also: Graduate School of Engineering, The University of Tokyo, Visiting Researcher

    Ryo Yoshida, Professor

Also: The Graduate University for Advanced Studies, Professor

RIKEN, Team Director

Zhejiang University

 Tianqi Deng, Assistant Professor

 Haoming Zhang, PhD student

University of Notre Dame

 Tengfei Luo, Professor

 Zhihao Xu, PhD student

Nanyang Technological University

 Kedar Hippalgaonkar, Associate Professor

 Wei Nong, PhD student

Eurecat, Technology Centre of Catalonia

 Pol Torres, Senior Researcher

Carnegie Mellon University

 Alan J. H. McGaughey, Professor

Oak Ridge National Laboratory

 Lucas Lindsay, Senior R&D Staff

Purdue University

 Xiulin Ruan, Professor

University of South Carolina

 Ming Hu, Professor

Cornell University

 Zhiting Tian, Professor

Journal Information

Journal: npj Computational Materials

Title: Database and deep-learning scalability of anharmonic phonon properties by automated brute-force first-principles calculations

Authors: Masato Ohnishi*, Tianqi Deng, Pol Torres, Zhihao Xu, Terumasa Tadano, Haoming Zhang, Wei Nong, Masatoshi Hanai, Zeyu Wang, Michimasa Morita, Zhiting Tian, Ming Hu, Xiulin Ruan, Ryo Yoshida, Toyotaro Suzumura, Lucas Lindsay, Alan J. H. McGaughey, Tengfei Luo, Kedar Hippalgaonkar, Junichiro Shiomi*

DOI: 10.1038/s41524-026-02033-w

Funding

This research was partially supported by the Japan Science and Technology Agency (JST) under the CREST program “Innovation of Disordered Thermal Functional Materials through QR-Loop Large-Scale Continuous Space Exploration (JPMJCR21O2)” and “Science and Application of Spatially and Temporally Localized Nano Thermal Phenomena (JPMJCR19I2).”

In addition, computational resources were provided through the MEXT Advanced Research Infrastructure for Materials and Nanotechnology (ARIM) program and the HPCI System Research Projects (hp220151, jh230065, hp240194, hp250202). These resources were supplied by the Information Initiative Center, Hokkaido University; the Cyberscience Center, Tohoku University; the Information Technology Center, The University of Tokyo; the D3 Center, Osaka University; and the Academic Center for Computing and Media Studies, Kyoto University.

Glossary

(Note 1) Phonons

Quasiparticles representing the collective vibrations of atoms in a crystal from a quantum mechanical perspective. They are the primary carriers of heat in solids, and phonon scattering and lifetimes determine the thermal conductivity of materials.

(Note 2) First-principles calculations

Computational methods based on fundamental laws of quantum mechanics that predict material properties without relying on empirical parameters. Techniques such as density functional theory (DFT) are used to calculate electronic states and lattice vibrations.

(Note 3) Anharmonic phonon properties

Properties arising from interactions between phonons (anharmonic interactions). These are key factors that determine lattice thermal conductivity and other thermal transport properties, including phonon lifetimes and scattering processes.

(Note 4) Lattice thermal conductivity

A physical property that describes how effectively heat is transported by lattice vibrations (phonons) in a crystal. It is distinguished from thermal conductivity due to electrons.

(Note 5) Harmonic phonon properties

Properties obtained when atomic vibrations are treated within the harmonic approximation (i.e., expanding the interatomic potential up to the second order). These include fundamental vibrational characteristics such as phonon dispersion relations, vibrational modes, and heat capacity, but do not account for phonon–phonon interactions or scattering.

(Note 6) Graph neural networks

A machine learning approach that represents atoms and their bonding relationships in a crystal as a graph structure, enabling prediction of material properties from structural information. It has been widely used in recent years for materials property prediction.

(Note 7) Scaling law

An empirical relationship in which the performance of a machine learning model improves according to a certain rule as the amount of training data or model size increases. In this study, it was shown that the accuracy of AI models predicting thermal conductivity improves exponentially as the amount of phonon property data increases.

(Note 8) Materials Informatics

A research field that uses data science and machine learning to analyze and predict material properties, thereby accelerating the discovery and design of new materials.

(Note 9) Materials Project

An open materials database developed primarily by a research team in the United States. It provides first-principles calculation data, such as crystal structures and electronic properties, for hundreds of thousands of materials, and is widely used in materials research worldwide.

(Note 10) AI for Science

A research field that leverages artificial intelligence (AI) to accelerate scientific discovery. By utilizing machine learning and large-scale data analysis, it promotes new insights across various domains, including materials discovery, climate science, and life sciences.