The cell-fidelity programme is trying to forecast how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation. One input to its forecast law is the capacity factor: the charge the cell takes at cycle 2 divided by the charge it takes at cycle n. The last step forecast that factor from the electrode balance (lost lithium sliding the open-circuit window along the two electrodes' potential curves) plus a ten-cycle line in the charge overpotential η. It got within 4 to 8% at cycle 99. The leftover error opened late, because the line kept falling after the real η had levelled off.
What I did
I used the same 40 PyBaMM DFN designs on four parameter sets, with no new simulation. I wrote the plan and the four hypotheses down and committed them before computing anything. Then I swapped the η line for a saturating fit, η(l) = c + d·exp(−(l − L₂)/λ), fitted to cycles 2 to 10 over a fixed grid of decay lengths. I also asked the question that decides where the programme goes next: if the forecast used the measured capacity factor, read from the full 99-cycle run, how much better would it get?
What came out
| set | factor error, η line | factor error, saturating η | node lever | node lever + measured factor |
|---|---|---|---|---|
Ai2020 | 7.6% | 2.4% | 2.3% | 2.0% |
Chen2020 | 6.7% | 4.5% | 5.9% | 6.3% |
OKane2022 | 6.8% | 4.3% | 7.6% | 6.9% |
NCA_Kim2011 | 3.6% | 4.8% | 4.8% | 3.5% |
The first two columns are set-median errors in the cycle-99 capacity factor. The last two are set-median errors in cycle-99 lithium lost to plating. With the saturating fit, the factor forecast is within 5% on all four sets. The decisive number is the last column. Even the exact factor moves the forecast by less than a point on three of the four sets. On Chen2020 it makes it slightly worse.
So the capacity factor is no longer where the ten-cycle forecast goes wrong. With the saturating factor, the forecast beats the law's own plating lever on 22 of 40 designs, short of the 26 I had set as the bar. On the two LG M50 sets it is still 5 to 7% off. An earlier step showed that feeding in the measured plating overpotential brings that down to about 1.5%. What's left is the plating overpotential's own drift, which rises over the first cycles and then turns over.
Limits
Each set has one pool of 8 or 12 designs, and earlier steps had already seen all of them. Half of the fitted decay lengths sit at an end of the grid, so nine points don't pin the saturating fit down well. η is a symmetric cut-off shift standing in for the real charge overpotential. All of this is one PyBaMM version and modelled fade, not cells on a bench. The full numbers and the pre-registered verdicts are in the project's log (milestone 5zs).