Inverse Design of Si/Ge Superlattices for Targeted Phonon Filtering using a NEGF-Machine Learning Framework

This work presents a novel inverse design methodology that combines accurate non-equilibrium Green’s function (NEGF) simulations with machine learning to engineer Si/Ge superlattices with tailored phononic properties. A multi-layer perceptron architecture (MLP) is trained to predict both the integrated thermal conductivity κ and the full phonon transmission spectra T(ν) from geometric descriptors, using a hybrid feature representation that combines statistical descriptors and principal component analysis (PCA)-compressed profiles. The model achieves excellent predictive accuracy, with a coefficient of determination of R2 ∼ 0.99 for κ and a root mean square error of  ∼ 0.013 a.u. for T(ν). The trained model enables rapid screening of 10^5 different structures, identifying optimal configurations that achieve a 10.20% reduction in κ compared to the best configurations in the original dataset. As a proof of concept, the methodology is applied to design a superlattice that selectively suppresses phonon transmission in the 15-36.2 meV (3.63-8.75 THz) frequency window, the characteristic emission range of the EL2 recombination center in GaAs, thereby demonstrating frequency-targeted thermal management at the source. In this use case, the optimized structure achieves a 38.84% reduction in transmission within the targeted frequency window. The proposed workflow is general and can be adapted to any electronic device where specific phonon frequency windows contribute to localized heating, offering a versatile pathway for integrated, at-source thermal management.

Palabras clave: Si/Ge superlattice, Inverse design, Phononic engineering, Thermal conductivity, Stop-band filtering, Active Cooling