Combinatorial development of efficient thermoelectric half-Heusler materials using machine learning, DFT and high-throughput experiments

• Applicants: Dr. Johannes de Boor; Professor Dr.-Ing. Alfred Ludwig
• Subject Area: Computer-Aided Design of Materials and Simulation of Materials Behaviour from Atomic to Microscopic Scale ; Synthesis and Properties of Functional Materials ; Physical Chemistry of Solids and Surfaces, Material Characterisation
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 545636554

The demand for green energy has constantly risen with increasing energy consumption and the impending climate change. This is exacerbated by unabated fossil fuel depletion whereas 60% of primary energy is wasted as unused heat. Thermoelectric (TE) materials are a viable green energy alternative as they convert heat directly into electricity. However, limited performance but also lack of mechanical robustness and usage of expensive, scarce, and toxic elements in to-date TE materials critically limit their application. Half-Heusler (HH) materials, equiatomic ternary compounds with high mechanical strength, have shown promise for TE applications and offer a huge compositional space with cheap, abundant, and environmentally benign elements that allows to vary their TE performance. The proposal targets at evaluating the full potential of the HH compositional space including iso- and alioelectronic substitution towards highly efficient TE materials by a combination of artificial intelligence tools and high-throughput synthesis and characterization. Machine learning (ML) can facilitate this exploration by combinatorial search, collation and application of small experimental datasets. However, deciphering the structure–property relationships from ‘black-box’ ML models, Density Functional Theory (DFT) validation of ML-screened substitutional compounds, but also experimental techniques for fabrication and high-throughput characterization of material libraries, involving microstructure, to validate ML and DFT results are major challenges here. A multi-throng approach involving ML, DFT, and experiments shall address these. It covers development of physically interpretable ML models using small-dataset compliant symbolic regression- and symbolic distillation-based ML techniques in an active learning framework for predicting the TE properties of the entire HH chemical space. TE properties predicted by the ML high-throughput screening will be validated by advanced DFT calculations on phase stability, electronic and thermal properties. To fully exploit the compositional space irrespective of the dopant-bound carrier concentration of the TE material, a novel concept using the material quality factor instead of the TE figure of merit zT as a guiding target parameter for ML screening will be rated. Experimental screening of promising compositions by thin-film and bulk material libraries and high-throughput characterization as well as synthesis of selected compositions as homogeneous bulk samples along with microstructural analysis will feed back to the ML training and shall resolve discrepancies between ML, DFT, and experimental results. This approach will not only facilitate the exhaustive combinatorial development of efficient HH compounds but will also allow for future inverse design of TE materials due to gene features that shall be identified from physically interpretable ML descriptors.

• DFG Programme: Research Grants
• International Connection: France
• Partner Organisation: Agence Nationale de la Recherche / The French National Research Agency
• Cooperation Partner: Professor Dr. Philippe Jund