Composable Neural Emulators for Thermoelectric Generator Design

A team led by Takao Mori at Japan's National Institute for Materials Science (NIMS) published this work in Nature (vol. 652, pages 643–649, 2026). The paper introduces TEGNet, a neural network that predicts thermoelectric generator performance with greater than 99% accuracy while using only 0.01% of the compute time of commercial finite-element solvers.

Image: LI, Airan | SAMURAI - National Institute for Materials Science - LI, Airan

What TEGNet does

According to the Nature abstract via RePEc, TEGNet is architecturally general across material systems and supports flexible combinations of material-specific emulators. That modularity lets researchers mix and match components to explore diverse device architectures quickly, rather than re-running expensive simulations for each configuration.

The team validated the approach experimentally, optimising two device types:

• MgAgSb/Bi₀.₄Sb₁.₆Te₃ segmented generators — 9.3% conversion efficiency
• Mg₃Bi₁.₄Sb₀.₆–MgAgSb n–p paired generators — 8.7% efficiency

Both rank competitively against previously reported devices.

Why it matters

Thermoelectric generators turn waste heat into electricity, but designing them means searching a large space of materials and geometries. In an accompanying Nature News & Views piece titled "AI speeds up design of heat-driven power generators," Ady Suwardi described TEGNet on LinkedIn as a physics-informed neural network with a "plug-and-play" design that cuts computation by thousands of times versus finite-element methods.

Authors and access

Authors include Airan Li, Xinzhi Wu, Longquan Wang, Gang Wu, Jiankang Li, Zhao Hu, Xinyuan Wang and Takao Mori. The paper is open access via NIMS SAMURAI (DOI: 10.1038/s41586-026-10223-1).

Sources: RePEc/Nature abstract, NIMS SAMURAI, LinkedIn — Ady Suwardi