Multi-task attention for doped thermoelectric properties prediction

Abstract Improving the performance of thermoelectric (TE) materials is essential for their wider adoption in sustainable energy and cooling applications. Impurity doping is a common strategy for enhancing transport properties, yet synthesizing every possible TE composition is infeasible, and high-fidelity ab initio simulations remain computationally expensive. Machine learning offers a viable alternative by learning a cheaper surrogate model, but conventional approaches often struggle to capture the diverse doping effects due to compositional similarity between doped and undoped materials. To address this challenge, we adapt the existing Compositionally restricted attention-based Network ( CrabNet ) into a multi-task (MT) variant, MT CrabNet , which uses self-attention to implicitly learn composition-specific dopant effects. MT learning enables simultaneous predictions of seven TE transport properties, leveraging their interdependence to enhance performance. We also explore different strategies for encoding external variables such as temperature. Trained on an experimental dataset covering over 1400 unique compositions, MT CrabNet consistently achieves top performance, with significant improvements for electronic thermal conductivity and electrical conductivity. Compared to its single-task variant, MT CrabNet delivers up to 1.5× improvement in R 2 . This establishes MT CrabNet as a valuable tool for accelerating the screening and discovery of high-performance TE materials, with a design extensible to doped systems beyond the TE domain.We also share the code and dataset used in this work.