A new composable neural-network emulator can predict thermoelectric generator performance with greater than 99 % accuracy while using just 0.0 1% of the computation time required by conventional finite-element solvers.
Study: Composable neural emulators accelerate thermoelectric generator design. Image Credit: luchschenF/Shutterstock.com
Saving this for later? Download a PDF here.
Thermoelectric generators convert heat directly into electricity and are widely studied for waste-heat recovery. Their performance depends not only on intrinsic material properties, but also on how devices are designed and operated.
A material’s efficiency is often summarized by the dimensionless figure of merit, zT, which combines the Seebeck coefficient, electrical conductivity, and thermal conductivity. But high-zT materials alone do not guarantee high-performing generators. Device efficiency also depends on material compatibility, leg geometry, architecture, and boundary conditions.
Optimizing all of those variables usually requires repeated finite-element simulations of coupled thermo-electrical equations, a process that is reliable but slow and computationally expensive.
A Neural Emulator Instead Of Repeated Simulation
In the Nature study, the researchers introduce TEGNet, a composable neural-network emulator designed to act as a fast surrogate for finite-element simulation in thermoelectric-generator design. Rather than solving partial differential equations for every new geometry or operating condition, TEGNet learns the relationship between device inputs and key device-level outputs.
The model takes device dimensions and boundary conditions as inputs, including leg geometry, hot-side temperature, cold-side temperature, and applied current. It directly predicts voltage and cold-side heat flow, from which power output and conversion efficiency are derived.
By focusing on these intrinsic outputs instead of detailed internal field distributions, the approach simplifies training and supports fast inference.
How TEGNet Was Trained
To build the model, the researchers generated training data using high-fidelity COMSOL simulations. The dataset spanned variations in geometry, temperature conditions, and operating current, allowing the network to learn across a broad design space.
For the benchmarked MgAgSb case, 1,200 samples provided the best balance between data-generation cost and predictive accuracy.
TEGNet is a fully connected neural network with three hidden layers. Validation and test results showed coefficients of determination above 0.999 for key outputs, indicating that the emulator closely tracked the simulation results it was trained to reproduce.
Speed Without Losing Key Design Accuracy
Across a wide range of operating conditions, TEGNet reproduced thermoelectric-generator behavior with near-perfect agreement to COMSOL. The paper reports strong agreement not only for voltage and heat flow, but also for derived power output and conversion efficiency.
Speed is the clearest advantage. In the benchmark reported by the authors, COMSOL required about 2,237 seconds per material simulation, whereas TEGNet produced equivalent predictions in about 0.25 seconds. That is roughly 0.01 % of the computation time, making rapid screening of large design spaces practical.
Composable Design Broadens What The Model Can Do
A key contribution of the study is composability. Because TEGNet predicts intrinsic outputs for individual components, material-specific emulators can be combined to model more complex device architectures without rebuilding full simulation workflows each time.
The researchers used this framework to study segmented generators and n-p paired systems. For segmented devices, TEGNet identified optimal material-length ratios in MgAgSb/Bi0.4Sb1.6Te3 generators. The experimentally optimized segmented device achieved a conversion efficiency of 9.3 % at 593 K, placing it among the stronger reported results in this class of thermoelectric generator.
For n–p paired generators, the model rapidly optimized geometric parameters such as cross-sectional area ratio and leg height in Mg3Bi1.4Sb0.6–MgAgSb devices. The results showed that the best configuration did not simply correspond to equal p-type and n-type leg areas, highlighting how simulation-guided design can reveal non-obvious design choices. The experimentally tested n-p paired generator reached 8.7 % efficiency.
Practical Engineering Relevance
The study also addresses contact losses, a key issue in devices. The authors show that electrical and thermal contact resistances can be incorporated into the TEGNet framework without retraining the network, while still maintaining strong agreement with COMSOL. That makes the approach more relevant to practical engineering problems, where parasitic effects often shape final device performance.
The paper makes a strong case that neural-network emulators can accelerate thermoelectric-generator design at the device level while preserving the accuracy needed for engineering decisions. It also suggests a useful bridge from materials discovery to device implementation by making it easier to test how promising materials perform in realistic architectures.
At the same time, the authors are careful about the model’s current scope. TEGNet is designed for rapid exploration of steady-state device design space, not for detailed field-level analysis or for capturing nonlinear, transient, or stress-related effects. Those remain important areas for future work.
Overall, the study shows how AI can make thermoelectric-device design faster, more flexible, and easier to scale across materials and architectures, without losing the technical detail needed for serious engineering use.
Journal Reference
Li, A., et al. (2026). Composable neural emulators accelerate thermoelectric generator design. Nature, 652(8110), 643-649. DOI: 10.1038/s41586-026-10223-1
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.