Deep Learning of Multi-Component Cathode Electrochemistry from Experiments

Disordered rocksalt materials are the most promising earth-abundant cathode materials for Li-ion batteries, and as such can enable scaling of Li-ion energy storage to many TWh/year production. However, the computational modeling for DRX is difficult as such modern battery materials can contain a large number of elements with substantial site disorder.<br/><br/>Instead of approaching from <i>ab-initio</i> modeling, we will demonstrate a novel approach to the modeling and prediction of electrochemistry (discharge voltage profile) of DRX materials. We applied a deep neural network (DNN) trained directly on a large amount of experimental results. The DNN is trained with an end-to-end learning scheme, that includes the redox information appropriately regularized. The DNN can interpolate and make predictions for compounds that have not yet been tested, which can accelerate the exploration of DRX and other electrode materials.