Waste heat is everywhere: car engines, industrial machinery, kitchen appliances—even your own body. Some of that lost energy can be converted into electricity using thermoelectric generators: compact, solid-state devices that produce power directly from temperature differences without the need for spinning turbines or moving parts. But designing materials that make these systems efficient has long been an engineering slog, requiring slow simulations and painstaking experiments to identify combinations that conduct electricity while blocking heat. Now researchers in Japan have built an artificial intelligence tool that can design thermoelectric generators 10,000 times faster than conventional approaches. Prototypes built based on the tool’s recommendations performed on par with today’s leading thermoelectric devices, the study found. The research, reported 15 April in Nature, could boost a long-promised, but not widely adopted, clean-energy technology by dramatically accelerating the search for affordable materials and device designs that efficiently convert heat into electricity. Takao Mori, deputy director of the Research Center for Materials Nanoarchitectonics in Tsukuba, Japan, and his team conducted the research. “It’s a solid piece of work,” says Zhifeng Ren, director of the Texas Center for Superconductivity at the University of Houston, who was not involved in the study, “and points to the future role that AI will play in the design” of such technologies. Thermoelectric Generators Convert Waste Heat Thermoelectric generators have been around for decades, quietly powering spacecraft, supplying electricity to gas pipelines in isolated locations, and running remote sensors in places where changing batteries is impractical. But high costs and modest performance metrics have largely confined the devices to niche applications. Hopes of broader deployment in oil refineries, steel mills, and other heavy industries have yet to materialize, leaving enormous quantities of waste heat untapped. Large power plants typically rely instead on steam-driven systems that convert heat into electricity by boiling water to spin turbines. Those systems are highly efficient at large scales but require moving parts, maintenance, and relatively high operating temperatures that make them ill-suited for recovering heat from scattered or lower-temperature sources. Thermoelectric generators work better for those jobs. Their compact, solid-state design allows them to harvest smaller amounts of heat from surfaces such as engine exhaust pipes, factory boilers, server racks and high-performance electronics where conventional turbines would be impractical. But progress in thermoelectric generators (TEGs) has long been hamstrung by the slow, painstaking design process. That’s because it requires researchers to hunt for materials that can simultaneously conduct electricity efficiently while blocking the flow of heat. Finding this rare pairing is essential for harnessing the Seebeck effect, a phenomenon in which a temperature difference across two semiconductors drives an electric current. To achieve that, researchers often spend days or weeks evaluating a single configuration by sifting through possible designs using slow physics simulations. AI Speeds Design of Thermoelectric Generators The new AI-based approach dramatically speeds that search. Dubbed TEGNet, the publicly available tool is built on a neural-network framework trained to approximate the complex physics equations that describe heat flow and electrical transport in thermoelectric materials. Instead of repeatedly solving these equations from scratch, the model learns how materials behave and treats them as modular components that can be combined in many different ways. This allows researchers to rapidly screen thousands of potential device architectures and estimate their performance in milliseconds. “This speed enables exhaustive exploration of design parameters, uncovering optimal device configurations that might otherwise be overlooked,” wrote materials scientists Jing Cao, from Singapore’s Agency for Science, Technology and Research (A*STAR), and Ady Suwardi at Chinese University of Hong Kong, in a commentary published in Nature. To test the approach, Mori’s team used TEGNet to optimize two types of generator designs. One, known as a segmented unicouple, stacks multiple thermoelectric materials together so each operates most efficiently within a particular temperature range. The second pairs two complementary semiconductors, known as n-type and p-type materials, that produce electricity when heat flows across them. After scanning thousands of possible configurations, the AI identified device geometries predicted to deliver strong performance. The researchers then fabricated prototype generators using spark plasma sintering, a method that rapidly compresses powdered materials into dense solid components using pulses of electric current. Both designs achieved conversion efficiencies of about 9 percent under temperature conditions typical of industrial waste heat, where thermoelectric devices are most commonly deployed. That number might not sound spectacular. But any technology that converts heat into electricity faces a built-in ceiling on efficiency, determined by the temperature difference between its hot and cold sides—a fundamental thermodynamic constraint known as the Carnot limit. Within those bounds, the new designs from Mori and his colleagues rank among the better-performing thermoelectric generators reported for this temperature range. And when it comes to thermoelectrics, even modest gains can matter: small improvements in efficiency can determine whether recovering waste heat is economically worthwhile or not. AI Finds Cheaper Thermoelectric Materials Another limitation in thermoelectrics is the cost of materials and fabrication. The field has long depended on semiconductor material such as bismuth telluride, which contains relatively scarce tellurium and often requires carefully controlled crystal growth and microstructural alignment to achieve high performance. This increases manufacturing complexity and expense. By contrast, Mori says, some of the AI-designed devices identified by TEGNet can be made using simpler fabrication approaches and, in some cases, avoid bismuth telluride altogether. Although full details remain confidential because of ongoing industry collaborations, he says, preliminary cost estimates suggest the designs could move thermoelectric generators closer to economic viability for industrial waste heat applications. “From the estimated cost,” Mori says, “we can project an industrially competitive power generation cost for the first time in thermoelectric history.”