The cell-fidelity programme forecasts how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation. The last milestone built an error budget and found that on the OKane2022 LG M50 parameters most of the remaining 6.7% error comes from one input: the rise of the anode's open-circuit potential (OCP) at the nodes where plating happens.
The forecast gets that rise from a "node lever". It assumes every node in the anode loses lithium at the same rate, a uniform shift in stoichiometry, and reads the OCP curve at each node's starting point. An earlier step had found that this lever over-reads at the separator side of the anode, which is where most of the plating is.
The idea
If a node's local stoichiometry moves more slowly than the uniform shift, cycles 3 to 10 should already show it. So for each of the 20 anode nodes I fitted one gain on the shift, the number that best matches that node's measured OCP rise over those eight cycles. A gain of 1 is the old lever. The fit runs in OCP space so the graphite plateau's flat steps can't blow it up. Nothing past cycle 10 is read. The hypotheses were written down and committed before any gain was computed (LOG.md, 5zv).
What happened
On OKane2022 the gains came out well below 1, between 0.1 and 0.8, with a plating-weighted mean of 0.61. Over ten cycles the nodes really do lag the uniform shift. Carried out to cycle 99, though, that ratio over-corrects. The gain lever now under-reads the cycle-99 OCP rise by 14.8 mV, where the uniform lever was 5.9 mV off, and the lithium-loss forecast ends up 6.9% off against 6.7%. Both OKane2022 hypotheses are refuted.
On Chen2020, the other LG M50 set, the same method works. The forecast goes from 5.4% to 3.9%, which is as good as feeding it the measured OCP rise. NCA_Kim2011 improves too, from 3.0% to 2.3%. Across all 40 designs, 22 improve. That is short of the 26 the sign test needed.
| set | uniform lever | per-node gain | measured OCP (ceiling) |
|---|---|---|---|
Ai2020 | 2.3% | 2.5% | 3.5% |
Chen2020 | 5.4% | 3.9% | 3.9% |
OKane2022 | 6.7% | 6.9% | 3.5% |
NCA_Kim2011 | 3.0% | 2.3% | 1.0% |
Set-median error in cycle-99 lithium loss, at the physical exponent, line through cycles 2 to 10.
A lead, not a fix
I also reported a variant with gains on the four separator-side nodes only and a gain of 1 everywhere else. It takes OKane2022's set median to 4.1%, close to the 3.5% ceiling. But it improves only 6 of the 12 designs, it makes Chen2020 worse (6.9%), and it wasn't one of the hypotheses. It needs a pre-registered test on fresh designs before anyone calls it a fix.
The reading is that a node's lag is not a constant fraction of the shift. It is large early and relaxes later. A straight-line drift forecast failed the same way two milestones ago. The next step is a gain that relaxes as lithium is lost. For an actual design search the advice doesn't change: fit the forecast's line through 30 to 50 cycles instead of 10.
Everything here is modelled fade in one PyBaMM version (26.10.1.1), on 40 designs the programme has seen before, and the node OCPs are model internals, not anything you could measure on a real cell. Figure: nodegain.svg.