AI-designed Si-Ge super-networks are pushing the boundaries of thermal transport
By combining quantum simulations and artificial intelligence, researchers have shown that the local arrangement of layers in silicon/germanium superlattices plays a decisive role in controlling heat transport.
References:
Shubham Tyagi, Julian G. Fernandez, Anass Sebbar, Natalia Seoane, Antonio Garcia-Loureiro, Marc Bescond, Machine-learning discovery of extreme coherent thermal transport governed by motif-level order in Si-Ge superlattices, npj Computational Materials - Published: 04 June 2026.
DOI: https://www.nature.com/articles/s41524-026-02172-0 (article in open access)
At the nanometre scale, heat is mainly transported by phonons, the vibrational quanta of the crystal lattice. In superlattices consisting of an alternating stack of silicon (Si) and germanium (Ge) layers, these phonons can either propagate coherently as waves or be scattered at the interfaces between materials. Understanding and harnessing this competition is essential for developing more efficient devices, whether to improve the cooling of electronic components or to optimise thermoelectric materials.
This research was carried out in the following CNRS laboratory:
Institut des Matériaux, de Microélectronique et des Nanosciences de Provence (IM2NP, Aix-Marseille Université / CNRS)
A collaboration between french researchers and researchers from the University of Santiago de Compostela (Spain) has developed an original approach combining the non-equilibrium Green’s functions (NEGF) quantum method with machine learning. Using a convolutional neural network trained on thousands of simulations, the researchers were able to efficiently explore a design space comprising nearly 800,000 Si-Ge superlattices compatible with the experimental constraints. Thanks to this strategy combining artificial intelligence and quantum simulations, the researchers only needed to evaluate around 1,200 structures directly – less than 0.2 per cent of the total space of possible configurations – drastically reducing the total computational cost.
The results surprisingly show that the most efficient structures are not necessarily periodic ones. The researchers demonstrated that the local order of the layers, on the scale of a few elementary motifs, strongly controls thermal conduction. Some configurations identified by artificial intelligence exhibit thermal conductivity up to 35 per cent lower than the graded structures traditionally used to block heat, whilst others result in an increase of more than 30 per cent in thermal conductivity compared with the best-known periodic superlattices. This discovery thus challenges the widely accepted notion that global periodicity is the main factor governing coherent phonon transport. The research demonstrates, on the contrary, that carefully organised local patterns can promote or inhibit the propagation of lattice vibrations with remarkable efficiency.
Beyond the Si-Ge system studied here, this approach paves the way for the AI-assisted design of nanostructured materials with tailor-made thermal properties. These results could contribute to the development of new thermal management strategies for future electronic components, data centres and energy conversion devices. They have been published in the journal npj Computational Materials.
