Ten battery cycles can't hand a plating forecast its two missing inputs, at least not by straight lines

Extrapolated from cycles 2 to 10, the plating overpotential drift and charge-capacity factor keep little of the oracle's gain: 5.7 to 10.1% cycle-99 error.

The cell-fidelity programme forecasts the lithium a battery design loses to plating by cycle 99, using a ten-cycle simulation. The last milestone found that the forecast law is close to right on LG M50 cells if you give it two inputs measured in the full simulation: the part of the plating overpotential that isn't the anode's open-circuit potential, and the charge the cell actually delivers each cycle. With both, the error was 1.4 to 1.5%.

Both of those inputs were read up to cycle 99, though, so that was a diagnosis and not a forecast. This step asked the obvious follow-up. Can the first ten cycles predict those two inputs well enough to keep the gain?

The simplest test

For each of the same 40 PyBaMM DFN designs I fitted a straight line, through cycles 2 to 10, to each input as a function of lithium lost:

  • the overpotential drift (the plating overpotential's rise minus the OCP's rise), added to the node lever from milestone 5zn;
  • the capacity factor, which is the charge at cycle 2 divided by the charge at cycle n.

I then ran the same ten-cycle forecast with these lines in place of the measured inputs. The four hypotheses were committed before any of it ran.

What came out

Set-median error in cycle-99 lost lithium:

parameter setlaw's own levernode leverten-cycle lines (F)both measured (5zp)
Chen2020 (LG M50)8.2%5.9%5.7%1.4%
OKane2022 (LG M50)5.9%7.6%7.2%1.5%
NCA_Kim20112.2%4.8%10.1%1.1%
Ai20208.3%2.3%7.0%8.0%

The forecast beats the law's own lever on 15 of 40 designs. The pre-registered bar was 26, and 20 or fewer counted as refuted, so that hypothesis is refuted. On Chen2020 it closes about a third of the gap to the oracle. On OKane2022 it makes things worse.

Why straight lines fail

The capacity factor accelerates. At cycle 99 the ten-cycle line gives 1.20 to 1.59. The simulation gives 1.53 to 2.00. The line under-reads on every set, because the delivered charge falls faster in lithium lost after cycle 10 than before it. Putting that line into the law's own lever made LG M50 worse (8.2% to 15.0% on Chen2020). The law's own capacity term already rises faster than the line does.

The overpotential drift turns over. On LG M50 it grows over the first ten cycles, to about +0.3 to +0.6 mV, then reverses: by cycle 99 it is about −2.5 mV. A line through the origin can't follow that, so it forecasts the wrong sign. The drift is small next to the 30 mV OCP rise, but it sits inside an exponent with a factor of 25.3 per volt, so small errors still count. The plan named this risk in advance, and it happened.

What it means

The previous milestone's diagnosis still holds: on LG M50 the law's form is fine once it gets the right inputs. But those inputs aren't linear over the first ten cycles, so no short extrapolation will supply them. For a design search the practical advice hasn't changed: fit the law's line through 30 to 50 cycles.

The capacity factor is the bigger failure, and it is also the one with a physical model available. The electrode-balance map from milestone 5r predicts from the design alone how the voltage window shrinks as lithium is lost. Feeding the law that map's capacity curve, in place of a straight line, is the next step.

Limits: 40 designs, all seen in earlier milestones; the simplest possible forecasts, not the best; node-level overpotentials are model internals; one PyBaMM version; modelled fade, not a tested cell. Numbers, code and the pre-registration are in the programme's log and experiments/drift/.

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