Cross-Property Transfer Learning and Generative Crystal Design for Accelerated Discovery of Low-Thermal-Conductivity Materials

Abstract

Identifying crystalline materials with intrinsically low lattice thermal conductivity (κL) remains a major challenge in thermoelectric design due to complex phonon transport mechanisms and the limited availability of high-quality datasets. Although first-principles methods provide accurate κL predictions, their high computational cost limits large-scale screening. Here, we present a unified deep learning framework that combines graph neural networks (GNNs), cross-property transfer learning (TL), and transformer-based generative crystal modeling for accelerated κL prediction and materials discovery. Benchmarking CGCNN, DeeperGATGNN, and ALIGNN on a curated dataset of 3925 materials identifies ALIGNN as the best-performing model, emphasizing the importance of capturing local bonding environments, bond-angle correlations, and many-body interactions. Cross-property TL across six auxiliary properties further reveals that energy-based source tasks provide the most transferable latent representations for κL prediction. To explore materials beyond existing databases, the transformer-based generative model CrystaLLM is employed to generate 10,000 unique crystal structures. Subsequent thermodynamic and phonon stability screening identifies dynamically stable compounds, with 17 materials exhibiting ultralow κL ≤ 0.50 W m-1 K-1. Overall, this work demonstrates that integrating cross-property TL with generative crystal modeling provides a robust and computationally efficient pathway for overcoming data scarcity and accelerating the discovery of next-generation thermoelectric materials.