Interpretable AI unlocks new thermal material frontiers

A research team from Wuhan University has developed an innovative machine learning framework that accelerates the discovery of materials with tailored thermal properties. By combining interpretable deep learning with multiscale computational techniques, the team achieved highly accurate predictions of lattice thermal conductivity (LTC) while also shedding light on the underlying physical mechanisms. Their approach maintains the predictive power of traditional "black-box" models but adds the physical interpretability. As result, several high-performance materials for thermal management were identified. This advancement not only speeds up the development of efficient thermoelectric devices and thermal control systems but also provides deeper insights into the fundamental process of heat transfer at the atomic scale.

The calculation and/or the measurement of the thermal conductivity of materials is a fundamental challenge in materials science, essential for developing technologies in energy management, electronics cooling, and thermoelectric applications. Traditional experimental methods, while accurate, are often time-consuming and resource intensive. Whereas, theoretical approaches, such as solving the Boltzmann transport equation (BTE) based on density functional theory (DFT) or molecular dynamics (MD) simulations, provide valuable insights but are limited by their high computational costs, restricting their use in high-throughput material screening.

This has driven the pursuit of machine learning (ML) techniques that can rapidly predict thermal properties with reasonable accuracy. However, many ML models, especially complex black-box algorithms, often lack interpretability, making it difficult to understand the underlying physical mechanisms governing heat transfer in materials.

Addressing these limitations, a research team from Wuhan University presented a work that introduced an innovative, interpretable deep learning (DL) framework designed to efficiently and accurately predict the lattice thermal conductivity (LTC) of materials. By combining the strengths of DFT calculations, high-throughput screening, and physically interpretable ML models, this approach not only accelerates the discovery of new thermal materials but also provides deeper insights into the phonon transport mechanisms that influence heat conduction. This synergistic strategy bridges the gap between computational efficiency and physical understanding, paving the way for smarter, faster material design in thermal management and energy conversion technologies.

Through sensitivity analysis and symbolic regression, key parameters such as vibrational free energy and elastic bulk modulus were identified from ten classes of physical features, constructing an LTC prediction model that rivals the accuracy of “black-box” models while offering explicit physical interpretability. Furthermore, by integrating graph neural networks (GNNs) and unsupervised k-means clustering, a high-throughput screening pipeline of "qualitative preliminary screening → quantitative prediction → experimental validation" was established. Using this framework, the team identified four high-performance thermal management materials (including three thermal conductors and one thermal insulator) from thousands of candidates. Later, these predictions were rigorously validated using DFT and MD calculations, showing excellent agreement.

The implications of this research are far-reaching, laying the foundation for efficient discovery of materials with customized thermal properties, an essential step toward improving energy efficiency, promoting sustainable development, and driving technological innovation. These advancements open new avenues for the development of high-performance thermoelectric materials, cutting-edge electronic devices, and energy-efficient insulation systems, all of which contribute to lower energy consumption and reduced greenhouse gas emissions. Importantly, the use of interpretable AI models deepens our understanding of heat transfer at the atomic level, fostering new scientific insights and equipping policymakers with data-driven guidance for energy strategy and environmental protection. In essence, this work not only accelerates technological progress but also aligns with broader societal objectives of sustainability, economic advancement, and environmental responsibility.

The research has been recently published in the online edition of Materials Futures, a prominent international journal in the field of interdisciplinary materials science research.

Reference:  Yuxuan Zeng, Wei Cao, Yijing Zuo, Tan Peng, Yue Hou, Ling Miao, Ziyu Wang, Jing Shi. Accelerating the discovery of materials with expected thermal conductivity via a synergistic strategy of DFT and interpretable deep learning[J]. Materials Futures, 2025, 4(4): 045602. DOI: 10.1088/2752-5724/ae08d0