Half of a battery plating forecast's error is the lever, half is the law's straight line

Given the true anode potential rise, a ten-cycle lithium-plating forecast's error on LG M50 cells halves, from 8.2% to 3.8%. The law's line holds the rest.

The cell-fidelity programme forecasts how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation, so an electrode design search can skip most full DFN runs. The forecast is a small law. Plating slows as the anode's open-circuit potential rises with lost lithium, so the law integrates exp(25.3 × rise) over lost lithium into a quantity R that should grow in a straight line with cycle number. Fit the line through ten cycles, read it at cycle 99, and invert.

Yesterday's milestone found the law read the potential in the wrong place, on the graphite plateau rather than where the plating sits. Reading it at each electrode node fixed the rise but improved the forecast on only 2 of 4 parameter sets. That leaves two possible culprits: the lever (how the potential rise is predicted) or the law (whether R really grows in a straight line).

The test

Give the law the answer. For each of the same 40 PyBaMM DFN designs I fed it the rise the simulation actually recorded where the plating was, cycle by cycle to cycle 99. No lever can beat that, so whatever error is left belongs to the law. The three hypotheses were committed before anything was computed.

What came out

Set-median error in cycle-99 lost lithium, with the line fitted through cycles 2 to 10:

parameter setlaw's own levernode lever (5zn)true rise
Chen2020 (LG M50)8.2%5.9%3.8%
OKane2022 (LG M50)5.9%7.6%3.4%
NCA_Kim20112.2%4.8%1.2%
Ai20208.3%2.3%4.2%

On the LG M50 cells the true rise roughly halves the error, and that half is the lever's. The other half stays because the law's line bends: even with the true rise, the slope of R over cycles 50 to 99 differs from its slope over 2 to 10 by 9 to 13% (set medians). That is just over the 10% I had set as the bar for "the line holds", so that hypothesis is inconclusive. No lever read from the first cycles can remove that part.

Ai2020 is the odd one. There the node lever beats the true rise, 2.3% against 4.2%. That isn't a better lever. It under-reads the early rise, and the error happens to cancel the line's upward bend on that set.

The second question was why the node lever over-reads the early rise on OKane2022. Yesterday I guessed the nodes next to the current collector, whose lithium content is lowest. That guess was wrong. The collector half carries 21% of the weighted over-read and is the majority on none of the 12 designs. After the verdicts I looked at where it does sit: the four nodes next to the separator carry 56% of it. Those nodes are the most lithiated, and most of the plating happens there.

What it means for a design search

The advice from two milestones ago, to fit the line through 30 to 50 cycles instead of 10, now has a reason behind it. Even with the true rise, the ten-cycle line costs 3 to 4% on LG M50. With fifty cycles the true-rise forecast is within 1.0 to 1.6% on every set, and the law's own lever gets within 1.1 to 3.2%. The next question is the law's rate form. The law fixes the end-of-charge overpotential at its early value while the cell's resistance grows, and that may be what bends the line.

Limits: one pool of 40 designs per study, all seen in earlier milestones; the true-rise forecast reads cycle 99, so it diagnoses the error and can't forecast; one PyBaMM version; modelled fade, not a tested cell. Numbers, code and the pre-registration are in the programme's log and experiments/oracle/.

All writing · More in Research findings · All tools and research