FOURIER NEURAL OPERATOR SURROGATE OF LITHIUM-ION BATTERY MODELS

High-fidelity pseudo-two-dimensional (P2D) models are used in battery design and diagnostics but are too computationally expensive for real-time battery management systems or large-scale design studies. Fourier neural operators (FNOs) offer a promising alternative by learning solution operators directly on spatial fields rather than their vectorized forms. We develop an FNO surrogate for a P2D model that ingests applied current, temperature, and internal state fields over time and predicts the full spatiotemporal evolution of lithium concentrations and electric potentials in both electrolyte and active material across the cell. On in-distribution test trajectories, the FNO achieves low relative errors in all major physical fields, including those that drive overpotential and terminal voltage. Under moderate shifts in temperature and charge rate, the FNO maintains accurate voltage trajectories and controlled error growth, illustrating inherent robustness to modest out-of-distribution conditions. FNO's valuation requires only milliseconds per trajectory and is several orders of magnitude faster than the underlying P2D model, enabling real-time battery management, fast design-space exploration, and model-predictive control with physics-grade fidelity. For context, we benchmark against the multilayer perceptron flow-map surrogate, which exhibits higher errors and greater degradation under the same test scenarios.