AI-designed Si–Ge superlattices push the boundaries of thermal transport

Publication in npj Computational Materials!

Controlling heat flow in nanostructures is crucial for digital technologies, AI, and power electronics, but the vast number of possible architectures makes optimization difficult. By combining quantum simulations with artificial intelligence, we have demonstrated that the local arrangement of layers in silicon–germanium (Si–Ge) superlattices strongly influences thermal transport.

At the nanoscale, heat is carried mainly by phonons—vibrations of the crystal lattice. In alternating Si–Ge layers, phonons can propagate coherently as waves or be scattered at material interfaces. Controlling this competition could improve electronic cooling and thermoelectric materials.

Combining the non-equilibrium Green’s function method with machine learning we show that a convolutional neural network enabled them to explore nearly 800,000 experimentally feasible superlattices while directly evaluating only about 1,200 (less than 0.2% of the design space) greatly reducing computational costs.

The best-performing structures were not necessarily the most periodic. Instead, thermal conduction was largely determined by the local ordering of a few elementary layer patterns. AI-designed configurations achieved thermal conductivity up to 35% lower than conventional graded structures and more than 30% higher than the best-known periodic superlattices.

These findings challenge the view that global periodicity primarily governs coherent phonon transport. They show that carefully arranged local patterns can efficiently promote or suppress lattice vibrations, opening the way to AI-designed nanostructured materials with tailored thermal properties for electronics, data centers, and energy-conversion devices.

Fig. 1: (a) Schematic representation of the device geometry used to model the Si–Ge superlattices. The red and blue regions represent the hot and cold contacts, respectively. The green and beige spheres represent silicon (Si) and germanium (Ge) atoms, respectively. (b) The architectures identified by the neural networks exhibit a thermal conductivity up to 35% lower than that of the reference graded structures and up to 33% higher than that of the best periodic superlattices. These results reveal that the local arrangement of the layers, rather than global periodicity, controls coherent phonon transport.

Ref : S. Tyagi, J. G. Fernandez, A. Sebbar, N. Seoane, A. Garcia-Loureiro and M. Bescond, "Machine-Learning Discovery of Extreme Coherent Thermal Transport Governed by Motif-Level Order in Si-Ge Superlattices," npj Comput. Mater. (2026).

https://doi.org/10.1038/s41524-026-02172-0