Abstract
Physics-based models are essential for accurate State-of-Health (SoH) monitoring in lithium-ion batteries, yet their reliance on complex partial differential equations (PDEs) often renders them too computationally expensive for real-time applications. This work introduces a Physics-Informed Neural Network (PINN) surrogate for the Single-Particle Model (SPM) that achieves a $10^5$x speedup over traditional numerical solvers. We propose a novel multi-fidelity training hierarchy and a merged architecture to address the training instabilities typically encountered when modeling stiff electrochemical systems. Our results demonstrate that the PINN surrogate provides a robust, high-speed alternative for parameter inference and real-time battery management.
Introduction
Effective management of electric vehicle fleets and grid-scale energy storage requires precise knowledge of internal battery states. While the Single-Particle Model (SPM) offers a physically grounded representation of battery dynamics, its implementation involves solving coupled non-linear PDEs. Traditional numerical methods, such as finite difference or finite element solvers, are accurate but slow, particularly when thousands of iterations are required for Bayesian parameter inference or real-time health diagnostics.
Purely data-driven machine learning models offer speed but require vast amounts of labeled experimental data, which is costly to obtain. Physics-Informed Neural Networks (PINNs) bridge this gap by incorporating physical laws directly into the neural network's loss function. However, the "stiff" non-linearities of electrochemistry, such as Butler-Volmer kinetics, often lead to optimization failures in standard PINN frameworks. This paper details a robust implementation strategy to overcome these hurdles.
Methodology
The core of our approach is a PINN that maps temporal ($t$) and radial ($r$) coordinates to battery states, including lithium concentrations ($c_s$) and electrical potentials ($\phi$).
As shown in Figure 1, we utilize a Merged Architecture. Instead of training separate networks for each physical variable, the model shares initial hidden layers. This design forces the network to learn a shared latent representation of the coupled physics, improving generalization and training efficiency. This "coupling as a feature" approach significantly outperformed split architectures in our testing.
To ensure stability, we implemented a Multi-Fidelity Hierarchical Training scheme. Training begins on a "low-fidelity" version of the physics—a linearized approximation of the electrochemical equations. Once the network masters this simplified regime, it is transitioned to a "high-fidelity" stage where it learns the non-linear corrections required for the full SPM. Furthermore, we strictly enforce initial and boundary conditions using mathematical distance functions, ensuring the model remains physically consistent from the start of training.
Results
The PINN surrogate was benchmarked against synthetic data from a high-fidelity finite-difference solver. The model demonstrated exceptional performance across constant current discharges (1C to 3C) and dynamic Hybrid Electric Vehicle (HEV) drive cycles.
Figure 2 illustrates the training process for 23 different realizations. We utilized a two-stage optimization strategy: initial training with the ADAM optimizer to navigate the complex loss landscape, followed by L-BFGS for fine-tuning and rapid convergence to the global minimum. This approach significantly reduced the variability between training runs, a common issue in PINN literature, and ensured the model did not converge to trivial solutions.
Key performance metrics include:
- Computational Speed: Once trained, the PINN surrogate solves the SPM approximately 100,000 times faster than a traditional PDE solver.
- Zero-Data Accuracy: The model achieved high precision even when trained exclusively on physics residuals, without any experimental data, demonstrating the power of physics constraints.
- Robustness: The multi-fidelity approach effectively mitigated the numerical stiffness of the Butler-Volmer equations, leading to stable convergence across multiple initializations.
Conclusion
This work demonstrates that PINNs can serve as highly efficient surrogates for physics-based battery models. By employing a merged architecture and a curriculum-based multi-fidelity training approach, we successfully addressed the numerical stiffness of the SPM. While the current model excels at standard discharge cycles, future work (Part II) will extend these techniques to the more complex Pseudo-2D (P2D) model and explore parametric PINNs for broader generalization across different battery chemistries.
The implementation is open-sourced in the PINNSTRIPES repository.
Summary updated