A lithium-plating forecast trick, tested on 32 battery designs it never saw

Anode node-gain profiles learned on 40 simulated cells still help a ten-cycle plating forecast on 32 new ones, but by less, and with big single-design misses.

The cell-fidelity programme forecasts how much lithium a battery design will have lost by cycle 99, mostly to lithium plating, from a ten-cycle simulation. The last three steps found that the forecast's weak input is how fast the anode's open-circuit potential rises at each point through the electrode's thickness. Each point lags the uniform shift the forecast law assumes, by a factor I call its gain. Gains learned from other designs of the same chemistry helped. A straight line of gain against anode thickness helped most of all, cutting OKane2022's error from 6.7% to 2.7%.

That line had a problem: I found it after seeing the data, and it was judged on the same 40 designs that suggested it. So this step tested it on designs it had never seen.

What I did

I drew 8 fresh designs per parameter set (32 in all) from the same design pools, excluding every design used before. I ran each in PyBaMM's DFN model for 100 cycles with SEI growth and lithium plating on, recording the plating and the anode potential at each of 20 points through the electrode. Before running anything I committed the plan and four hypotheses. The main forecast was a gated thickness line: use the line only on a set where, in the training designs, the separator-side gain rises with thickness (Spearman correlation at least 0.5), and otherwise fall back to the plain uniform lever. Every profile was fitted on the old 40 designs only and read nothing of a new design past cycle 10.

What came out

Set-median error in the cycle-99 lithium-loss forecast, on the 32 new designs:

setuniform levermedian profile3 nearestthickness linemeasured-OCP ceiling
Ai20204.4%5.4%5.6%5.0%3.6%
Chen20207.5%7.5%7.0%6.5%4.1%
OKane20227.2%6.2%6.7%5.6%1.5%
NCA_Kim20113.3%2.8%2.9%2.5%0.3%

Three things held up. Every borrowed profile reads the cycle-99 potential rise better than the uniform lever on all four sets. On NCA_Kim2011 the thickness line cuts that potential error fourfold, from 19.9 to 4.9 mV. And the thickness line has the lowest median forecast error of any profile on three of four sets.

Two things did not. The gains are smaller than they were in-sample, and they are lumpy: the line beats the uniform lever on only 19 of 32 designs. On OKane2022 it takes four designs from 6 to 8% error down to 1 to 4%, but misses two thin ones by 14 to 15%. Those two sit at or below the thinnest training design, where a line in thickness has to extrapolate. And my pre-registered gate was the weak link. Its 0.5 threshold was a guess, and OKane2022 came in just under it at 0.46. So the gated forecast fired only on NCA_Kim2011 and threw away OKane2022's largest gain. Of the four hypotheses, two were supported (the gated set improves; the median profile replicates on OKane2022) and two were inconclusive.

One side lesson about reproducibility. Rounding the simulation records to seven significant digits, as an earlier step did to save space, moved Chen2020's results by up to 0.6 points. The forecast takes differences of cumulative lithium totals, and rounding eats those differences. I re-ran the bench at full precision. The re-run reproduced the first report character for character, and that full-precision file is the one committed.

What it means for a design search

The advice barely changes. To trust a plating forecast, fit its line through 30 to 50 cycles. On a chemistry like NCA_Kim2011, where the anode's gain plainly rises with thickness, a thickness-line profile from a few fully simulated designs gets you there in ten cycles. Elsewhere it improves the typical design but can badly miss a single one, so it is not a rule yet.

Next: use the line only inside the thickness range it was trained on, and test that rule on another fresh draw.

Code, data and the full log: research/cell-fidelity (milestone 5zy in LOG.md; figure experiments/freshnode/freshnode.svg). These are model results from one PyBaMM version, not measured cells.

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