The cell-fidelity programme forecasts how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation. This morning's step found a floor under that forecast on LG M50 cells. Even with every input to the fade law measured, the law's own ten-cycle line misses cycle 99 by about 2% at the median. The cause was not a shared curve that a longer fit or a borrowed bend could fix: each design's rate drifts a few per cent up or down over the run, and the direction varies by design.
So the next question was whether that drift can be seen early. If ten cycles show which way a design will drift, the forecast can correct for it, or a design search can at least flag the designs it should not trust.
What we tested
We defined a drift multiplier m for each of 104 PyBaMM DFN designs: how far the law's rate over cycles 10 to 99 departs from its ten-cycle slope (m = 1 means the line was right). Before computing any correlation we wrote down three ten-cycle observables: where in the anode the plating sits at cycle 2, how that position moves by cycle 10, and how lithium loss itself bends over the first ten cycles. We also fixed four hypotheses and the correction: a straight-line fit of m on whichever observable correlates best with m on the other designs, leaving the design being forecast out.
What we found
The drift is predictable. On both LG M50 parameter sets the plating depth at cycle 2 correlates with m at Spearman ρ = 0.80 (permutation p = 0.0003 after a Holm correction). Plating that sits nearer the separator means the rate speeds up later. The correlation holds inside each of the three design pools, so it is not an artefact of how the pools were drawn.
It does not lower the median. With the correction the set-median cycle-99 error is 2.3% and 2.1%, against the line's 1.8% and 1.9%. The typical design drifts by about 4%, and that is about the size of the correction's own error, so there is nothing to gain in the middle of the set. That hypothesis is refuted.
It does fix the tail. We only looked at quantiles after the pre-registered report, so this part is a post hoc finding. The drift is heavy-tailed: a quarter of the designs drift by 10 to 25%, and those are the designs the plating depth flags. The correction cuts the upper-quartile error from 8.0% to 3.5% on Chen2020 and from 7.3% to 3.0% on OKane2022. The 90th percentile drops from 10.1% to 6.5% and from 9.0% to 5.3%.
On the real ten-cycle forecast, the one that guesses its inputs instead of measuring them, the same correction beats the line on all four parameter sets. The biggest gain is on NCA_Kim2011, where the median error goes from 4.0% to 1.4%.
Why it matters
A design search loses most from a few badly wrong forecasts, not from a 2% error everywhere. This result says those forecasts can be found before cycle 11: they are the designs whose early plating sits far from the set's typical depth. They can be corrected, or sent to a longer simulation. The programme's next step is that decision rule, scored on DFN calls saved at a target tail error.
Limits: 24 to 28 designs per parameter set; the plating depth is a model internal that a real cell shows only through a diagnosis; one law form; one PyBaMM version; modelled fade only. Numbers are in experiments/drift7c/results.txt and the post hoc tail check is in posthoc.txt.