Using neurocomputing techniques to determine microstructural properties in a Li-ion battery

Current ab-initio approaches such as Quantum Mechanics (QM) calculations or Molecular Dynamics (MD) simulations to study the doped cathode structures are computationally expensive. In this work, we present the development and application of neural computing models to study the crystal structure of the cathode materials in Lithium-ion batteries, Lithium Manganese Oxide (LMO) in particular. We do this using LMO crystal configurations doped with Aluminum. We successfully demonstrate the application of 8 multi-layer perceptron models that are capable of predicting the potential energy of LMO crystal configurations, with coefficients of determination (R2$$R^2$$) ranging from 0.95 to 0.98. To achieve this, models were developed by training and testing them on the potential energy (eV) values of over 460,000 crystal configurations. In lithium-ion battery research, the developed Neural Network models could be utilized alongside existing atomic or molecular simulation tools to efficiently identify optimal crystal configurations that could be subjected to more detailed investigation. With the integration of the multi-layer perceptron models of this work, the total time to evaluate all possible crystal configurations can be reduced by approximately 88% than when using just QM and MD simulations for such evaluations.