Thermoelectrics and Waste Heat Recovery

Thermoelectric materials convert heat gradients directly into electricity, offering a pathway to recover the enormous amounts of waste heat lost in industrial processes, vehicle exhaust, and power generation. However, achieving high thermoelectric performance requires simultaneously optimizing three interrelated transport properties — electrical conductivity, Seebeck coefficient, and thermal conductivity — that are physically coupled in ways that make independent optimization impossible. The figure of merit ZT remains below 2.5 for most known materials at practical operating temperatures, and the materials that do achieve high ZT often rely on expensive, toxic, or scarce elements. The field needs a generative approach that navigates the coupled optimization landscape rather than screening individual properties sequentially.

Traditional thermoelectric material development uses Boltzmann transport theory and DFT band structure calculations to evaluate ZT for known or proposed compositions, followed by experimental optimization of carrier concentration through doping studies. This evaluate-then-optimize approach is inherently limited: it can only assess candidates that researchers propose based on chemical intuition, and the coupled nature of thermoelectric transport properties means that improving one metric (e.g., reducing thermal conductivity through nanostructuring) often degrades another (e.g., carrier mobility). ML models trained on thermoelectric databases reproduce known structure-property correlations but cannot generate compositions in unexplored regions where the conventional trade-offs between transport properties might be circumvented.