AI Finds Materials That Barely Conduct Heat express
Heat flows through matter like water through pipes: sometimes fast, sometimes barely. Engineers seek materials that almost don't let heat pass — like clogged pipes. Such insulators are vital for energy saving and devices that turn heat into electricity. The neural network KappaFormer quickly evaluates the "throughput" of an atomic structure.
KappaFormer's secret lies in its physical intuition. Ludwig Boltzmann realized that heat depends on the chaos in particle motion. In a crystal, atoms quiver; synchronized vibrations carry heat, chaotic ones don't. The network distinguishes these contributions by analyzing the atomic pattern. It was trained on millions of data points about the compression and stretching of materials, and then fine-tuned on rare examples of thermal conductivity. Amazingly, this was enough for ultra-precise predictions — like a plumber detects a clog by the sound of the pipe.
After checking thousands of options, KappaFormer found substances with near-zero heat leakage — hundreds of times better than window glass. Their atomic geometry entangles thermal vibrations, and they die out like sound in an untuned piano. Now the network is being retargeted to search for other materials. The physics stitched into the algorithm leads to materials of the future — from space shields to quantum computers.
Original abstract
as written by the authors arXiv:2604.03547
Machine learning has been widely used for predicting material properties. However, efficient prediction of lattice thermal conductivity ($$\kappa_\mathrm{L}$$) remains a long-standing challenge, primarily due to the scarcity of high-quality training data. Here we introduce KappaFormer, a physics-aware Transformer architecture that embeds the harmonic-anharmonic decomposition of $$\kappa_\mathrm{L}$$ within the network. KappaFormer comprises a harmonic branch pre-trained on large-scale elastic property data and an anharmonic branch fine-tuned on limited experimental $$\kappa_\mathrm{L}$$ data, enabling effective knowledge transfer and enhanced generalization. High-throughput screening with KappaFormer identifies multiple candidates with ultralow $$\kappa_\mathrm{L}$$, which are further confirmed by first-principles calculations. Physics interpretability further elucidates the vibrational mechanisms governing thermal transport suppression, linking structural motifs to strong anharmonicity. This study provides a generalizable framework for physics-guided machine learning to accelerate the discovery of new materials.
Among the discovered materials, some hold heat hundreds of times more reliably than double-glazed windows — essentially, an almost eternal thermos.
In Dune, the heroes' stillsuits act as perfect heat insulators — today, with the help of AI, scientists are approaching the creation of similar materials.