Predicting Li-Ion Migration Energy Barriers in Battery Cathode Materials via Convolutional Neural Network Model Based...

Predicting the Li-ion migration energy barrier in battery cathode materials via machine learning has attracted increasing attention. However, the capture of complex material features via the accurate definition of structural descriptors and cross-scale information integration for neural network modeling remains challenging. Benefiting from the reported divide-and-conquer strategy, we propose a convolutional neural network (CNN) model employing hybrid geometric-topological descriptors for the efficient prediction of Li-ion migration energy barriers, in which the geometric descriptors capture local Li-O polyhedra of initial and transition states, as well as the topological descriptors derived from persistent homology characterize structural connectivity and ring channels along the migration pathway. Compared with the recurrent neural network (RNN) and Fourier-feature network (FFN) models, the CNN model optimized via residual block structures and L2 regularization achieves a mean absolute error (MAE) of 0.0589 eV. The minimum Li-O distance during the Li-ion migration is identified as the most critical factor affecting the migration barrier, and the comparable importance scores of d_min, d_std, and H0 bars suggest a synergistic effect of multiple structural descriptors, highlighting the necessity of adopting hybrid geometric-topological descriptors. The present work provides an efficient and accurate approach for high-throughput screening of materials with rapid Li-ion diffusion.