Machine-learning discovery of extreme coherent thermal transport governed by motif-level order in Si-Ge superlattices

Controlling phonon-mediated heat transport in multilayer nanostructures is central to thermal management and energy-conversion technologies. Yet identifying architectures exhibiting extreme lattice thermal conductivity (κl) remains challenging due to the combinatorial complexity of interface arrangements and the high computational cost of transport simulations. In superlattices, thermal transport arises from competing wave effects: coherent phonon propagation enhances transport, whereas interface scattering and localization suppress it. We study Si-Ge superlattices using an iterative machine-learning-guided non-equilibrium Green’s function (NEGF) approach, where a convolutional neural network trained on NEGF data efficiently explores an experimentally constrained design space to identify both low- and high-κl limits. Dense interface clustering induces cumulative elastic scattering and reduces κl by up to 35%, while sparse, locally ordered motifs preserve extended phonon modes and enhance κl by up to 33% beyond periodic designs. These results are obtained at 200 K in a regime where phonon transport is predominantly coherent and elastic; thus, the identified structural trends are representative of the coherent transport regime. Our work reveals that local motif-level order, rather than global periodicity, governs extreme thermal transport in multilayer systems.