Prediction and optimization of ionic conductivity in nanoparticle-based electrolytes using convolutional neural networks

We demonstrate the application of deep learning approaches for prediction and optimization of nanocomposite properties. Our recent study on nanoparticle-based electrolytes suggested the nanoparticle configuration in oligomeric solvent host to be a key design parameter influencing ionic conductivity. Navigating the vast structural search space for identifying optimal nanoparticle microstructures at a fixed nanoparticle loading, is however limited by the computational cost of the accompanying simulations. To overcome such challenges, in this work, we develop a convolutional neural network (CNN) model for quantitative prediction of ionic conductivities. The model is trained using selected mesoscale simulation results. The structure-property linkages established using the CNN model exhibit better predictive capability compared to that using deep learning models based on physics-inspired approaches. By integrating the trained CNN models with a topology optimization algorithm, we demonstrate accelerated morphological space search to identify nanoparticle networks with enhanced ionic conductivities.