Incorporating accurate artificial-intelligence (AI) models into high-throughput workflows is needed to accelerate the discovery of new high-performing thermoelectric materials. However, the scarcity of available data, for the thermoelectric figure of merit, ZT, and the components that comprise it, as well as the significant effort required to acquire new data limits their creation. Our goal is to solve this problem by developing AI-guided, high-throughput workflows to calculate the thermoelectric properties of a material.
Recently we developed high-throughput workflows for calculating the lattice dynamics of a material to all orders of anharmonicity in FHI-vibes. Using these workflows we are then able to generate a dataset to learn models for a material thermal conductivity using SISSO++, and find heuristics that describe where in materials space new thermal insulators can be found.
We now look to extend these workflows by introducing active learning into them. First we will generate a large database of candidate materials, including all of their relevant vibrational properties. From here we will use acquisition functions to automatically rank these materials by their likelihood for being good thermal insulators or containing new information for the models. Finally we will calculate the thermal conductivity of the most promising candidates using the ab initio Green Kubo method and update the models and ranking.
While currently the workflows are generated for only the thermal conductivity of a material, it is an additional goal of ours to extend this functionality to all aspects of a materials thermoelectric figure of merit, ZT. This work will be done in collaboration with Christian Carbogno as the various pieces are ready to be implemented into FHI-vibes.
Software and method development play an integral part of our research efforts. As a result we will be maintaining and expanding upon the FHI-vibes and SISSO++ codebases, as well as contributing to other projects.