Data-driven surrogate modelling and calibration of particle breakage in Li-ion electrode calendering

This study introduces a comprehensive, data-driven framework for efficient simulation and calibration of particle breakage during lithium-ion battery electrode calendering. We first proposed a Discrete Element Methode (DEM) based spherical damage model realised via a particular particle replacement mechanism, and generated a broad training set by uniformly varying eight model parameters. On this basis, a novel dual-branch neural network is trained to reproduce both the full compaction-pressure profile and the statistical distribution of fracture events, achieving high fidelity against held-out simulations. This surrogate model is embedded within a two-stage inverse-optimisation routine that infers breakage parameters directly from experimental pressure-depth data, yielding calibrated inputs for rapid forward prediction. With the help of surrogate efficiency, a global Sobol analysis quantifies the relative importance of each parameter, guiding model simplification work and highlighting the roles of characteristic size and fracturing fitting exponent. The coupled DEM-Multilayer Perceptron (MLP) studies reveal that allowing particle breakage during heavy calendering appreciably lowers tortuosity and enhances effective diffusivity compared to non-fracturing scenarios, with larger initial particle sizes further improving transport pathways. Collectively, this work delivers a scalable and advanced approach for parameter estimation, sensitivity assessment, and design optimisation of electrode microstructures under realistic processing conditions.