High-Throughput Thermoelectric Materials Screening by Novel Deep Convolutional Neural Network with Fused Orbital Field Matrix and Composition Descriptors

Thermoelectric materials assist in generating or harvesting energy through converting waste heat into reusable electricity. Thermoelectric materials are also widely used in real-world applications, such as refrigerators and cooling electronics. However, most popular and known thermoelectric materials to date were proposed and found by intuition, mostly through experiments. Unfortunately, it is extremely time and resource consuming to synthesize and measure the thermoelectric properties through trial and error experiments. Here, we develop a convolutional neural network (CNN) classification model that utilizes the fused orbital field matrix (OFM) and composition descriptors to screen a large pool of materials to discover new thermoelectric candidates with power factor higher than 10 μW/cmK². The model used a few thousand data generated by density functional theory (DFT) coupled with AMSET, a package for calculating electronic transport properties that does not assume constant relaxation time for electrons, thus ensuring more reliable electronic transport properties calculations. The classification model was also compared to traditional machine learning algorithms, such as light gradient boosting and random forest. With the classification model we screened >35,000 structures with non-zero band gap from the Open Quantum Materials Database (OQMD). We identified lots of potential high performance thermoelectric materials with ZT > 1 across wide temperature range for both n- and p-type doping with different doping concentrations, including quaternary Heuslers, half-Heuslers, and more material families. Insight into the correlation between high power factor and some representative material descriptors was also gained by feature importance analysis and maximal information coefficient, which provides new simple routes for fast screening promising thermoelectrics from large-scale material pool in the future.