A battery plating-forecast fix that failed its second held-out test

A thickness-line correction that helped a ten-cycle lithium-plating forecast on 32 new battery designs did not hold on 32 more. Here is why.

Yesterday's step in the cell-fidelity programme tested a correction to a ten-cycle forecast of how much lithium a battery design loses by cycle 99, mostly to lithium plating. The correction is a straight line, per point through the anode's thickness, of how much that point's open-circuit potential lags the uniform shift the forecast assumes. On 32 designs it had never seen, it gave the lowest median error on three of four parameter sets. But its two worst misses sat at the thin end of the training range, where a line extrapolates.

So today's step tried the obvious repair and tested it on another 32 fresh designs. The line was fitted on all 72 designs simulated so far, not

  1. It was used only when a new design's anode thickness fell inside the

training range, with the plain uniform lever as the fallback otherwise. The plan and four hypotheses were committed before any new simulation ran.

What came out

Set-median error in the cycle-99 lithium-loss forecast, on the 32 new PyBaMM DFN designs (8 per set, 100 cycles each):

setuniform leverline, 40 designsline, 72 designsline in rangemeasured-OCP lever
Ai20201.2%2.6%3.4%3.4%8.5%
Chen20204.5%5.4%4.7%4.7%7.0%
OKane20225.1%6.5%6.2%6.2%4.6%
NCA_Kim20115.3%4.8%4.7%3.7%0.7%

The repair did not work. The range-restricted line beats the uniform lever on one set of four, NCA_Kim2011. Design by design it wins 14 of the 29 designs inside the range, a coin flip. Over both fresh pools the line wins 33 of 64. Three of the four hypotheses were refuted. The one that held is that more training designs make the line itself a bit better (three sets of four), and that was not enough.

Why

The last column is the telling one. It replaces the uniform lever with each design's own measured anode potential rise, which no ten-cycle forecast could know. If the lever were the forecast's main error, that column would be the floor. On this pool it beats the uniform lever on only 14 of 32 designs. On Ai2020 and Chen2020 it is worse at the median. So on three of the four sets, getting the lever right does not fix the forecast: other terms dominate, and errors in the lever cancel them as often as they add. NCA_Kim2011 is the exception on every pool so far. There the lever is the error (0.7% with the measured lever), and a borrowed profile helps.

The previous step's gains on three sets were, it now seems, mostly the luck of one draw of eight designs per set. That is exactly what a second pre-registered draw is for.

What it means for a design search

Do not correct a ten-cycle plating forecast with a node-gain profile borrowed from other designs, except perhaps on cells like NCA_Kim2011 where the anode's potential rise is the dominant error. Fit the forecast's line through 30 to 50 cycles instead. This closes the node-profile thread of the programme (five steps). Next is the error budget on all 104 designs simulated with through-thickness records, to name what does limit the forecast on the other three sets.

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

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