Designing materials for thermoelectric applications: density functional theory and machine learning

Abstract: Thermoelectric devices, which can convert heat into electricity, have the potential to significantly enhance future green energy systems, if materials with optimal electron and phonon transport characteristics become available. In my talk I will present computational strategies combining DFT and machine learning (ML) for the investigation of thermoelectric materials. In addition to prototype chalcopyrite, CuFeS2 [1, 2], we have studied a wide range of chalcopyrite-structured chalcogenides [3,4] and pnictide compounds [5]. While the electronic transport properties of these materials are attractive, they suffer from too high thermal conductivities. To afford accurate predictions across this large family of compounds, we solve the phonon Boltzmann transport equation with force constants derived from DFT and ML-based regression algorithms, reducing by about two orders of magnitude the computational cost with respect to conventional approaches of the same accuracy. The results allow us to rationalise the role of chemical composition, temperature, and nanostructuring in the thermal conductivities, and to predict interesting compositions for thermoelectric applications within this important family of semiconductors. I will also show how machine learning techniques can be employed in a more data-intensive approach to identify promising compositions for thermoelectric applications within a wider chemical space [6]. We have developed a neural network model with Transformer architecture which can predict electron transport coefficients for a given temperature and doping level, from knowledge of the material’s composition [7]. A webapp is available for easy interaction with the model [8].

[1] Tippireddy et al. Chemistry of Materials 34 (2022) 5860.
[2] Tippireddy et al. Journal of Materials Chemistry A10 (2022) 23874.
[3] Plata et al. Chemistry of Materials 34 (2022) 2833–2841.
[4] Plata et al. Journal of Materials Chemistry A11 (2023) 16734.
[5] Posligua et al. ACS Applied Electronic Materials (2023). In press.
[6] Antunes et al. in Machine Learning in Materials Informatics: Methods and Applications 2022, ACS Publications.
[7] Antunes et al. Machine Learning: Science and Technology 4 (2023) 015037.
[8] https://thermopower.materialis.ai/

Bio: Dr Ricardo Grau-Crespo grew up in Cuba, where he studied Physics at the University of Havana. He completed his PhD in Computational Materials Science at Birkbeck, University of London. He is currently an Associate Professor of Materials Theory at the University of Reading, where his group uses a range of computational techniques, from first principles to machine learning, in the investigation of energy materials, mainly for thermoelectric and photocatalytic applications.