AI slashes the time needed to design better heat-harvesting devices

 

From wearable technology to industrial heat recovery, thermoelectric generators which convert waste heat into electricity have an enormous range of potential applications. So far, however, designing high-performing versions of these devices has remained a painstaking task.

Now, through new research published in Nature, Airan Li and colleagues at the National Institute for Materials Science in Japan have developed an AI-based tool that predicts device performance with greater than 99% accuracy, all while cutting computational time by around 10,000-fold.

The optimization problem

Thermoelectric generators produce electricity when one side is hotter than the other: a temperature difference drives electric charge through the device, generating a current. Yet beyond this fairly simple concept, getting the process to work efficiently is surprisingly tricky. It isn't enough to simply choose materials with good thermoelectric properties: the materials must also be compatible with each other, meaning their ability to transport heat and electrical charge must be well-matched.

On top of that, the length and cross-sectional area of each component need to be carefully tuned, since these dimensions directly control how much electrical resistance and heat flow the device produces. For researchers, the process of finding optimal combinations of materials, geometries, and architectures through physical experiments alone would be prohibitively slow.

For this kind of problem, researchers would normally turn to computer simulations, allowing them to test thousands of designs virtually. However, these simulations need to solve complex equations describing heat and electrical flow simultaneously. Running enough of them to thoroughly explore the entire design space of thermoelectric generators can take days, weeks, or even months.

Introducing: TEGNet

To sidestep this bottleneck, Li's team developed a new machine learning tool named TEGNet, which they trained using data generated from conventional simulations. Rather than repeatedly solving it from scratch, this approach allowed the model to learn the underlying physics, resulting in a system that could make a single performance prediction in just a few milliseconds, rather than tens of minutes.

A particular strength of TEGNet is its modularity. For their demonstration, the researchers built separate models for individual thermoelectric materials, which could then be snapped together like building blocks to simulate more complex device architectures. These included designs that stack multiple materials in series, or pair materials with different charge-carrier types.

Elegantly optimized designs

Using this approach, TEGNet was able to scan thousands of configurations, and identify optimal geometries for two experimental prototypes. Respectively, these devices converted 9.3% and 8.7% of the heat energy passing through them into useful electrical energy—both ranking among the best reported conversion efficiencies for their operating temperature range.

The team's approach points toward a broader shift in how thermoelectric devices are designed. By compressing design cycles from months to minutes, tools like TEGNet could open up the field to smaller research groups without access to powerful computing infrastructure. In turn, it could help to accelerate the journey from promising materials to practical, real-world energy-harvesting devices.