The cell-fidelity programme asks how much lithium a battery design will lose to plating by cycle 99, using only a ten-cycle simulation. After four milestones today, the capacity factor is solved to within a few per cent. The error that is left on LG M50 cells (5.4 to 6.7%) sits in one input: the part of the plating overpotential that does not come from the anode's open-circuit potential. I'll call it the drift.
The idea
Over cycles 2 to 10 the drift rises by a fraction of a millivolt while the effective charge overpotential falls and starts to level off. Milestone 5zs had already fitted a saturating curve to that fall. So the forecast for this step ties the drift to it: drift = k × (how far the charge overpotential has fallen), with one slope k fitted on the ten cycles. The ten-cycle straight line from milestone 5zq is the comparison. Four hypotheses were committed before anything was computed. The plan also named the risk: a tie to a quantity that falls and levels off can rise and level off, but it can never turn over.
What came out
On 40 PyBaMM DFN designs across four parameter sets:
| set | drift error at cycle 99, tie | line | forecast error, no drift term | with the tie | with the measured drift |
|---|---|---|---|---|---|
Ai2020 | 3.5 mV | 13.0 mV | 2.3% | 4.5% | 4.9% |
Chen2020 (LG M50) | 6.0 mV | 8.7 mV | 5.4% | 7.6% | 2.4% |
OKane2022 (LG M50) | 4.9 mV | 7.1 mV | 6.7% | 7.9% | 6.5% |
NCA_Kim2011 | 0.02 mV | 0.02 mV | 3.0% | 3.0% | 3.0% |
The tie is a better drift forecast on every set, and a worse plating forecast on every set where the drift is not zero. On LG M50 the drift turns negative after cycle 10 (about −2.5 mV at cycle 99). Every ten-cycle forecast I tried keeps it positive, so adding any of them pushes the plating forecast the wrong way. Refitted over the whole run, the tie's slope changes sign on both LG M50 sets. Within ten cycles nothing hints at the turn-over.
The last column is a ceiling: the node lever given the true drift. It splits the two LG M50 sets. On Chen2020 the true drift halves the error, so the drift really is the missing piece there. On OKane2022 it barely helps, so something else is behind that set's 6.5%. On Ai2020 the true drift makes the forecast worse, which means the node lever's other term already accounts for it.
What it means
Cutting a forecast input's error is not the same as cutting the forecast's error. Here the sign of the input mattered more than its size. Of the ten-cycle options, leaving the drift out or freezing it at its cycle-10 value works best. For a real design search the advice from earlier milestones still holds: fit the forecast line through 30 to 50 cycles, not
- The next step is to find out what turns the drift over on
Chen2020.
Limits: the same 40 designs every milestone has used (a check on known designs, not a held-out test); a one-slope tie; the drift is a model internal; one PyBaMM version; modelled fade only. Plan, outcome and numbers: LOG.md, 5zt.