NIMS developed TEGNet (Thermoelectric Generator Neural Network), a neural network for designing thermoelectric generators by utilizing artificial intelligence (AI). TEGNet can predict performance of a power generator, a process which used to take enormous computational time with traditional simulation techniques, with only about 1/10,000 of the time conventionally needed, while maintaining over 99% accuracy. This technology significantly accelerates optimization from material development to device design, and is expected to be applied to waste heat recovery and stand-alone power supply for IoT sensors, for example. This research result was published in Nature at 11:00 U.S. Eastern Standard Time, April 15, 2026 (0:00 Japan Standard Time, April 16, 2026).
Toward realizing a sustainable society, thermoelectric generation technology capable of generating power indefinitely merely by installing a device at any place with temperature differences is drawing attention. In order to improve performance of thermoelectric generators, not only material development, but also optimal design of dimensions and structure are indispensable. However, conventional numerical analysis (finite element method) had a problem of requiring computation to be repeated whenever conditions were changed, which caused heavy computational load and made large-scale and high-speed design exploration difficult.
In order to solve this problem, the research group developed TEGNet, an AI model capable of optimizing the design of thermoelectric generators at high speed. If material properties and element dimensions and conditions are input into TEGNet, TEGNet quickly predicts voltage and heat flow that are generated within the device, making it possible to estimate power generation output and conversion efficiency at high accuracy (Figure 1). The most notable feature of this technology is in its composability, which allows designers to freely combine independent TEGNet models that have been trained for each material, like blocks, based on the laws of physics. This approach can enable designers to speedily and exhaustively explore and optimize performance, even for a device with a complex structure that combines materials having different properties. To demonstrate this approach, the team optimized two types of device designs using Mg-Sb (magnesium-antimony) based materials, prototyped and evaluated those devices, and achieved high conversion efficiencies of up to 9.3% and 8.7% under practical temperature conditions.
This research result proposes a next-generation design technique using AI as core technology, as opposed to conventional thermoelectric generator design dependent on numerical simulations. While many AI studies have focused on material-level optimization in recent years, the key feature of this research is that it directly targets device-level optimization. As a result, it becomes possible to sophisticate material design and device design in a complementary manner, and AI utilization is expected to develop further not only in the thermoelectric field, but in the entire energy field.
Nature
Experimental study
Not applicable
Composable neural emulators accelerate thermoelectric generator design
15-Apr-2026