Researchers at POSTECH and UNIST used topology optimization to design non intuitive thermoelectric leg geometries (asymmetric hourglass and I beam shapes) that achieved 8.2x the power per unit area of conventional rectangular designs when fabricated in a bismuth antimony…
The computer-designed geometry looked like it should not work. An asymmetric hourglass, an I-beam with a narrow waist — shapes that would draw the eye past their obvious wrongness without explaining why they were drawn that way. The POSTECH and UNIST researchers who fabricated these geometries in a bismuth antimony telluride alloy using 3D printing, then ran them under real convective heating and cooling conditions, found they produced eight times more electrical power per unit area than the standard rectangular thermoelectric modules found in car exhaust systems and industrial facilities. The lab results matched the computer model's predictions almost exactly. That agreement — the simulation-to-hardware translation that kills most papers in this field — is what separates this result, as described in their paper in Nature Communications.
Thermoelectric generators convert heat directly into electricity with no moving parts, which makes them mechanically simple and reliable. The problem is the geometry. Maximizing power output requires optimizing simultaneously for heat flow through the hot side, electrical resistance on the cold side, and mechanical stress across a wide temperature range. Rectangular legs are easy to manufacture and have dominated thermoelectric module design for decades. They are not what the physics wants. Topology optimization — a computational method that distributes material across a design domain to maximize a specified objective function — ran all of these constraints together and found shapes no engineer would have proposed, producing the I-shaped and asymmetric hourglass geometries described in Nature Communications.
The best design, under the strongest thermal boundary conditions, produced 8.2 times the power per unit area of a conventional rectangular generator, per Nature Communications. Under other tested conditions, the improvement ranged from 79 percent upward. The 8.2x figure is a best-case result, not a specification — the paper reports a family of geometries each optimized for its thermal environment.
The limitation is that this is a single thermoelectric leg, not a full generator. Scaling up means printing larger areas with uniform grain structure, managing thermal expansion mismatches between materials, and integrating the leg architecture with headers, cold plates, and electrical contacts. The paper does not address these steps. Large-scale implementation, fabrication complexity, and cost remain open for the field, as noted in a ScienceDirect review of next-generation thermoelectric generators. A steel mill or trucking fleet wanting custom thermoelectric harvesters today would find the manufacturing path unclear and the per-watt cost uncompetitive.
What the paper demonstrates is that the bottleneck in waste heat recovery was never the material. It was the shape. And the shape, it turns out, is a solved problem — the remaining work is engineering and economics.
That conclusion is supported by a pattern that spans decades and industries. NASA used genetic algorithms in 2006 to evolve an antenna for the ST-5 spacecraft that looked like a bent paperclip and outperformed every human-designed candidate. Aerospace has produced topology-optimized structural members that initially looked wrong to engineers trained on conventional designs. The POSTECH result is the latest instance of the same empirical finding: when the optimization landscape is large enough and the objective function well-defined, the machine finds solutions human intuition would have rejected.
The difference now is manufacturing. Topology optimization has been generating non-intuitive geometries for decades in structural engineering. The constraint was always fabrication — complex shapes that the computer specified could not be manufactured at scale with conventional methods. 3D printing has changed that constraint. The POSTECH team used extrusion-based 3D printing, a process now accessible in most research environments, to produce the optimized legs at the device level. The shapes the computer wanted and the shapes the factory could make are no longer in conflict.
The 8.2x is not the real number. The real number is zero — the distance between what the computer predicted and what the hardware delivered. That gap has killed more thermoelectric papers than it has advanced. Closing it means the playbook for designing custom harvesters for any waste heat stream — automobile exhaust, industrial furnaces, semiconductor plants, data center cooling systems — now exists in validated form. The remaining question is who closes the engineering and economics gaps, and how fast.
The researchers are Professor Jae Sung Son of POSTECH's Department of Chemical Engineering and Professor Hayoung Chung of UNIST's Department of Mechanical Engineering. Their paper appeared in Nature Communications on February 19, 2026.