Our energy-cell-dfn study found that the fast single-particle model with electrolyte (SPMe) and the full Doyle-Fuller-Newman model (DFN) disagree badly on thick, low-porosity electrode designs at high rate, and traced the disagreement to the electrolyte running out of salt. The open question was which model a real cell agrees with. So we ran both, plus the plain single-particle model, against the measured discharge curves that PyBaMM distributes with two published parameter sets: Ecker et al. (2015), a Kokam 7.5 Ah cell at 1C and 5C, and Ai et al. (2020), an Enertech 2.28 Ah pouch cell at 0.1C to 2C.
What came out (voltage RMSE against the measurement, isothermal):
- At low rate the data cannot separate the two models. At 0.1C and 1C they differ by under 1 mV, and each misses the measurement by 40 to 52 mV. That error is in the open-circuit curve and the end-of-discharge knee, not in the electrolyte physics.
- At 2C on the Ai cell, DFN wins clearly: 18 mV against 53 mV. This is the case where the electrolyte drops furthest, to 30% of its starting salt. The gap between the models grows as the salt runs down: 0.4, 3, 10 and 41 mV at 97, 84, 67 and 30% of the initial concentration.
- At 5C on the Ecker cell, neither model fits (66 and 71 mV), and SPMe is slightly closer. There, the thermal model changes the error by 26 mV and the mesh by 15 mV, both more than the choice of model.
A practical note for anyone running these comparisons: PyBaMM's default mesh was not converged at 5C. The error moved from 86 to 71 mV between the default mesh and four times as many points, while every lower-rate case moved by under 1 mV.
What this does not show: neither measured case reaches the near-empty electrolyte where SPMe failed on our design candidates, and neither cell is the LG M50 cell those designs are based on. It supports reading the full model rather than the reduced one in that region; it does not validate either model there. The script, data hashes, tests and full table are in the project directory, and the hypotheses were written down before the run.
Produced with AI assistance (Claude).