Machine Learning-Enhanced Multiscale Computational Framework for Optimizing Thermoelectric Performance in Nanostructured Materials

The direct conversion of solid-state heat to electricity using thermoelectric materials has attracted attention; however, their effective application is limited because of the challenge of ensuring a balance between the microstructural features at the quantum, mesoscale, and continuum scales. Current computational and machine-learning methods have a small design space, wherein few to no interactions between the electronic structure, phonon transport, and device-level are considered. This makes it difficult to discover stable high-figure of merit (ZT) settings that are manufacturable and strong in the actual working environment. This study presents a multiscale hybrid optimization framework that combines first-principles descriptors, synthetic microstructure optimization, machine-learning surrogate modeling, Finite Element Method (FEM)-based transport modeling and optimization, and an uncertainty-sensitive reinforcement-learning optimization framework. The results of the performance improvements are compared with those of physics-only, ML-only, and recent optimization baselines using hybrid thermoelectric. The integrated framework offers an accuracy of 97.2%–95.8% in predicting the transport coefficients and offering 18%–32% ZT improvements from the baselines. The optimized configurations remained stable under ±10% fabrication-style perturbations, confirming that the discovered designs were not fragile numerical artifacts. The proposed approach provides a valuable solution for finding a reliable way to obtain high-ZT, fabrication-tolerant thermoelectric designs, which opens the way to accelerated material discovery and the design of next-generation thermoelectric (TE) devices.