PINK: physical-informed machine learning for lattice thermal conductivity

Yujie Liu, Xiaoying Wang, Yuzhou Hao, Xuejie Li, Jun Sun, Turab Lookman, Xiangdong Ding, Zhibin GaoJournal of Materials Informatics 2025 · DOI

TL;DR

This paper introduces PINK, a novel framework that uses physical-informed machine learning to rapidly predict lattice thermal conductivity (κL) directly from crystal structure files. By combining a crystal graph convolutional neural network (CGCNN) with a simplified physical model, PINK achieves high accuracy and scalability, accelerating the discovery of materials for thermal management and energy conversion.

What Problem Does This Paper Address?

Efficient thermal management is critical for modern electronics and energy conversion technologies, and lattice thermal conductivity (κL) is a key property governing heat transport. Traditional methods for predicting κL, such as solving the phonon Boltzmann transport equation or performing molecular dynamics simulations, are computationally expensive and time-consuming, hindering large-scale material screening. While empirical models like the Slack model offer faster alternatives, they still require significant effort to obtain crucial parameters like sound velocity and the Grüneisen parameter. This creates a bottleneck in discovering new materials with desired thermal properties.

Methods

The core of the PINK framework lies in its hybrid approach. It leverages a Crystal Graph Convolutional Neural Network (CGCNN), a powerful deep learning architecture for materials science, to predict essential material properties: the bulk modulus (B) and shear modulus (G). These predictions are made directly from crystallographic information files (CIFs), which contain the atomic structure and bonding information.

Once B and G are predicted, PINK employs a simplified physical formula (derived from the Slack model) to estimate κL. This formula requires parameters like the average sound velocity and the Grüneisen parameter, which are then derived from the predicted B and G, along with other crystal information extracted from the CIF. This "physical-informed" aspect ensures that the machine learning predictions are grounded in physical principles, leading to more reliable and interpretable results. The entire process is automated and accessible through a user-friendly web application.

Datasets

  • Training/Validation/Testing for CGCNN: The paper mentions using the Matbench V0.1 dataset, specifically the "matbench_log_gvrh" and "matbench_log_kvrh" subsets, which contain elastic modulus data for approximately 10,987 materials.
  • κL Prediction: For predicting κL, datasets were sourced from the AFLOW database and relevant publications. A test dataset was constructed using crystal structures, Grüneisen parameters, and κL values from AFLOW and literature.
  • High-Throughput Screening: The primary dataset for screening comprised 377,221 stable materials from the Materials Discovery Database (as reported by Merchant et al.). This dataset was then filtered based on band gap, stability, and the absence of radioactive elements.

Key Insights and Ideas

  • Bridging ML and Physics: The paper demonstrates a successful synergy between data-driven machine learning (CGCNN) and physics-based models (Slack approximation) to predict a complex material property like κL. This hybrid approach offers the best of both worlds: the speed and scalability of ML with the interpretability and physical grounding of established theories.
  • CIFs as a Universal Input: The ability to predict κL directly from CIF files is a significant advancement. CIFs are widely available and contain rich structural information, making this approach broadly applicable without requiring extensive pre-processing or experimental data.
  • Accelerated Material Discovery: By drastically reducing the time and computational cost of predicting κL, PINK enables rapid screening of vast material spaces. This is crucial for identifying promising candidates for applications like thermoelectrics, where ultralow κL is highly desirable.
  • Identification of Novel Candidates: The screening process successfully identified promising materials with ultralow κL, such as Ag₃Te₄W and Ag₃Te₄Ta, which are then validated through first-principles calculations.

Key Results

  • Accuracy: The CGCNN model for predicting bulk and shear moduli achieved a Mean Absolute Error (MAE) below 13 GPa and an R² value approaching 1 on the test dataset. For κL predictions, PINK achieved an MAE of 0.442 and an R² of 0.653 when compared to DFT-calculated values, and an MAE of 0.526 and an R² of 0.881 when compared to experimental values.
  • Speed: PINK can predict κL and related properties in seconds, a dramatic improvement over traditional methods that can take hours to weeks.
  • Scalability: The framework was applied to a dataset of 377,221 materials, identifying 11,869 candidates with ultralow κL values (< 1 W·m⁻¹·K⁻¹).
  • Validation: First-principles calculations confirmed the ultralow κL values for Ag₃Te₄W and Ag₃Te₄Ta, validating the PINK predictions.

Limitations and Future Work

  • Grüneisen Parameter Accuracy: The paper notes that the accuracy of κL predictions could be significantly enhanced with more precise Grüneisen parameters. While the current method estimates it, more accurate experimental or first-principles data for this parameter would be beneficial.
  • Model Generalization: While the CGCNN model shows good generalization, its performance is contingent on the quality of its elastic modulus predictions. Further improvements in CGCNN architectures or training data could enhance overall accuracy.
  • Future Directions: The authors suggest integrating PINK with experimental databases and materials informatics platforms to further accelerate discovery. They also propose combining it with tools for optimizing electrical transport properties to design better thermoelectric materials. Additionally, exploring advanced CGCNN architectures (e.g., OGCNN, MEGNet, GeoCGNN) and more sophisticated descriptors (e.g., SOAP, Voronoi tessellations) could lead to even higher prediction accuracy. The precise engineering of materials considering mechanical properties and environmental stability alongside thermal conductivity is also highlighted as a future avenue.

Summary updated

Abstract

Lattice thermal conductivity (κL ) is crucial for efficient thermal management in electronics and energy conversion technologies. Traditional methods for predicting κL are often computationally expensive, limiting their scalability for large-scale material screening. Empirical models, such as the Slack model, offer faster alternatives but require time-consuming calculations for key parameters such as sound velocity and the Grüneisen parameter. This work presents a high-throughput framework, physical-informed kappa (PINK), which combines the predictive power of crystal graph convolutional neural networks (CGCNNs) with the physical interpretability of the Slack model to predict κL directly from crystallographic information files (CIFs). Unlike previous approaches, PINK enables rapid, batch predictions by extracting material properties such as bulk and shear modulus from CIFs using a well-trained CGCNN model. These properties are then used to compute the necessary parameters for κL calculation through a simplified physical formula. PINK was applied to a dataset of 377,221 stable materials, enabling the efficient identification of promising candidates with ultralow κL values, such as Ag3Te4W and Ag3Te4Ta. The platform, accessible via a user-friendly interface, offers an unprecedented combination of speed, accuracy, and scalability, significantly accelerating material discovery for thermal management and energy conversion applications.
BibTeX
@article{liu2025pink,
  author = {Yujie Liu and Xiaoying Wang and Yuzhou Hao and Xuejie Li and Jun Sun and Turab Lookman and Xiangdong Ding and Zhibin Gao},
  title = {{PINK: physical-informed machine learning for lattice thermal conductivity}},
  journal = {Journal of Materials Informatics},
  year = {2025},
  doi = {10.20517/jmi.2024.86},
}
Figure