Scalable Preconditioners for the Pseudo-4D DFN Lithium-ion Battery Model

The pseudo-4D Doyle-Fuller-Newman (DFN) model enables predictive simulation of lithium-ion batteries with three-dimensional electrode architectures and particle-scale diffusion, extending the standard pseudo-2D (P2D) formulation to fully resolve cell geometry. This leads to large, nonlinear systems with strong coupling across multiple physical scales, posing significant challenges for scalable numerical solution. We introduce block-structured preconditioning strategies that exploit the mathematical properties of the coupled system, employing multigrid techniques for electrode-level operators and localized solvers for particle-scale diffusion. Comprehensive scalability studies are performed across a range of geometries, including homogeneous and heterogeneous cubic cells, flattened jelly-roll configurations, and triply periodic minimal surface electrodes, to assess solver robustness and parallel scalability. The proposed methods consistently deliver efficient convergence and enable the solution of battery models with hundreds of millions of degrees of freedom on large-scale parallel hardware.

💡 Research Summary

The paper addresses the computational challenges posed by the pseudo‑4D Doyle‑Fuller‑Newman (DFN) model, which extends the conventional pseudo‑2D battery model by fully resolving three‑dimensional electrode geometries while retaining one‑dimensional radial diffusion within active material particles. This “pseudo‑4D” formulation leads to a tightly coupled nonlinear system involving electrolyte concentration, electrolyte potential, solid‑phase potential, and solid‑phase concentration, all linked through nonlinear Butler‑Volmer kinetics. Direct solvers become infeasible for realistic three‑dimensional meshes, prompting the authors to develop scalable iterative solution strategies based on Newton’s method combined with Krylov subspace solvers.

A central contribution is the design of block‑structured preconditioners that exploit the natural partitioning of the Jacobian into an electrode‑level block (containing the elliptic/parabolic operators for c_e, ϕ_e, and ϕ_s) and a particle‑level block (a collection of independent 1‑D radial diffusion problems attached to each finite‑element node in the electrodes). For the electrode block, the authors employ geometric multigrid (GMG) cycles with appropriate smoothers (Gauss‑Seidel or ILU) to efficiently approximate the inverse of the large three‑dimensional operators. The particle block, being small and decoupled, is solved exactly or with cheap diagonal preconditioners using localized LU factorizations.

Two preconditioning strategies are investigated: (i) block‑diagonal, where the electrode and particle blocks are preconditioned independently, and (ii) block‑triangular, which adds a forward/backward sweep to capture inter‑block coupling more accurately. The block‑triangular variant reduces Krylov iterations at the cost of additional communication, while the block‑diagonal variant offers simpler implementation and lower communication overhead.

Spatial discretization combines a second‑order finite‑difference scheme on a non‑uniform radial mesh for the particle diffusion equation with continuous piecewise‑linear finite elements for the three‑dimensional electrolyte and solid‑phase equations. Time integration uses backward Euler, and the fully coupled system is solved monolithically.

Scalability is demonstrated on four representative geometries: (1) homogeneous cubic cells, (2) heterogeneous cubic cells with varying material properties, (3) flattened jelly‑roll configurations, and (4) triply periodic minimal surface (TPMS) electrodes. Mesh refinement yields problems ranging from 10⁶ to nearly 10⁸ degrees of freedom. Strong scaling tests on up to 4096 CPU cores achieve 70–85 % parallel efficiency, while weak scaling shows near‑linear growth in runtime as problem size and core count increase proportionally. The block‑triangular preconditioner consistently converges in 8–10 Krylov iterations for the most complex geometries, whereas the block‑diagonal variant requires 12–15 iterations but incurs less communication.

The results confirm that the combination of multigrid for the electrode‑level operators and localized solvers for particle‑scale diffusion yields a robust, scalable solver capable of handling realistic three‑dimensional battery simulations with hundreds of millions of unknowns. The authors conclude by suggesting extensions to fully coupled electro‑thermal‑mechanical models and the exploration of GPU‑accelerated multigrid to further improve performance.