KappaFormer: Physics-aware Transformer for Lattice Thermal Conductivity via Cross-domain Transfer Learning
Abstract
A physics-aware Transformer combines harmonic and anharmonic branches with cross-domain transfer learning to predict lattice thermal conductivity from limited high-quality data.
Type
Status
Open access
KappaFormer embeds a harmonic-anharmonic decomposition into its architecture and transfers knowledge from large-scale elastic-property data to smaller thermal-conductivity datasets.

Authors
AI for Science · Computational Chemistry · Computer Vision
Yu Zhu works at the intersection of AI for Science, computational chemistry,
and computer vision. His research spans machine learning force fields,
physics-driven reconstruction of molecular orbitals from scanning tunneling
microscopy images, open-set fine-grained recognition, and materials modeling.