The cell-fidelity programme asks when a lithium-ion electrode design search can trust PyBaMM's cheap SPMe model instead of the full DFN model. Its trust test is the electrolyte's salt. If a design's minimum salt concentration drops below 10% of where it started, it gets checked with DFN. The last step found something awkward: the best design in a pool keeps landing right at that threshold. On 21 (cell, current) conditions the best design's closed-form salt number, r_kt, sat between −0.06 and +0.58, while the pools ranged from −3.98 to +0.92. So the designs the cheap model gets wrong are the same ones a search is looking for. This step asked why.
The score splits three ways. Every milestone ranks designs by DFN's Wh/kg at cycle 100. That number is exactly the product of three factors:
- U, the cell's thermodynamic energy per kg. PyBaMM computes it from the parameter set without simulating anything.
- Utilisation, the share of U that the first cycle actually delivers at this current.
- Retention, the share of that left at cycle 100.
U and r_kt both come from the parameter set and the design alone. Together they cost 0.3 s per design.
What we predicted, then ran. We split the score on the 21 conditions already simulated, and wrote down four predictions before running anything new. Then we ran 389 new simulations: 64 fresh designs, each run through SPMe and DFN, on three conditions picked by r_kt alone. Those were Ecker2015 at 1.5 times the usual current, Mohtat2020 at 3 times and Ai2020 at 0.375 times. All four predictions held:
- Among designs that don't starve their electrolyte, U ranks first-cycle energy at Spearman 0.99, 1.00 and 1.00.
- Among designs that do, r_kt ranks utilisation at 0.82, 0.60 and 0.88. The
Mohtat2020figure sits exactly at the bar we set. - Across a pool, U and r_kt pull against each other: −0.82, −0.81 and −0.36. Thicker, denser electrodes store more energy and starve their electrolyte sooner.
- A shortlist built from U and r_kt alone (their Pareto front, with no SPMe run) gets within 0.9%, 4.7% and 2.7% of the best design. It needs 11 to 18 DFN runs out of 64.
So the optimum sits where the first limit binds. Above the salt flag utilisation hardly moves (0.85 to 0.96), and U decides the ranking. Below it utilisation falls with r_kt. The best design is the highest-energy one before the salt runs out, which puts it just above the edge: r_kt +0.22 on Ecker2015 and +0.12 on Mohtat2020.
Ai2020 is the exception, and it fits the same picture. On that cell, designs with more energy fade faster: U and 100-cycle retention correlate at −0.94. So fade is the limit that binds first. The best design sits well clear of the edge (r_kt +0.67) and in the bottom third by energy. It is the first of 24 conditions where the best fell outside the range the last step predicted.
What it costs to skip SPMe. The SPMe-based procedures from earlier milestones found the exact best on all three pools. They used 15 to 40 DFN runs, plus SPMe on most designs. The front gives up a few percent and runs no SPMe at all. The part it misses is retention, which is the one factor with no closed form here.
For an electrode designer: treat r_kt ≈ 0 as a constraint, maximise thermodynamic energy up to it, and spend full-model runs on fade. That is three conditions, one PyBaMM version and stack-level numbers, not a tested cell. In development, NCA_Kim2011 at half current showed that U alone can mislead where other losses are large. Its utilisation was 0.80 even above the flag.
Plan, verdicts and caveats: research/cell-fidelity/LOG.md. Code and data: research/cell-fidelity/experiments/edge/.