Lithium plating sits off the graphite plateau, and a battery degradation forecast was reading the plateau

Read at each node's own stoichiometry, a plating law's lever matches the simulated OCP rise to 2% on LG M50; the forecast gains on 2 of 4 cells.

The cell-fidelity programme is trying to forecast how much lithium a battery design loses to plating over 100 cycles from a short simulation, so an electrode design search does not have to run the full DFN model for every candidate. The forecast is a small law: plating slows as the anode's open-circuit potential rises with lost lithium, at a rate set by an exponent. Physics says that exponent is about 25 per volt (αF/RT). On LG M50 cells the law needed about 35.

The last two milestones narrowed it down. The exponent the law needs is 25 times the ratio of the anode potential rise where the plating happens to the rise the law itself computes. On LG M50 the law under-reads the rise by about 1.4 times. And early in life the gap is far bigger: by cycle 10 the potential where the plating sits has risen 2.9 mV on Chen2020, while the law says 0.4.

The guess

The law reads the anode potential at the top of its charge window, assuming the whole electrode sits at one lithium content. On graphite, the top of that window is on a flat plateau, where losing lithium barely moves the potential. If the plating happens somewhere else on the curve, the law is reading the wrong part of it.

So this milestone kept the law's own account of how lost lithium shifts the lithium content, but read the potential at each of the 20 electrode nodes' own lithium content after cycle 2 (inverted from the simulated node potential), weighted by where the plating was. Nothing in that lever uses more than cycle 2, so in principle a ten-cycle forecast can use it. The five hypotheses were committed before the lever was computed.

What came out

The guess about the rise holds up. On LG M50 and Ai2020 the plating sits 0.15 to 0.22 higher in lithium content than the law's window top, off the plateau. Read there, the law's shift reproduces the measured plating-weighted rise:

cell setmeasured / node lever, cycles 3 to 99measured / law's own lever
Ai20201.00about 1.0
Chen2020 (LG M50)0.981.49
OKane2022 (LG M50)0.841.35
NCA_Kim20111.181.04

On Chen2020 the node lever predicts 3.25 mV of rise by cycle 10 against 2.88 mV simulated and the law's 0.42.

The forecast does not follow everywhere. At the physical exponent, with only ten cycles, the cycle-99 lithium-loss error moves:

cell setlaw's levernode leverdesigns improved
Ai20208.3%2.3%7 of 8
Chen20208.2%5.9%9 of 12
OKane20225.9%7.6%5 of 12
NCA_Kim20112.2%4.8%1 of 8

That is 22 of 40 designs, no better than a coin. All five pre-registered hypotheses came out inconclusive: each was met on two sets and missed on the other two.

What it means

The 1.4 times factor on LG M50 is mostly a matter of where the potential is read, not different plating physics. That is worth knowing for anyone building a reduced plating model on graphite: a single, uniform lithium content at the top of charge hides the plateau edge where the plating actually is.

But a lever that gets the rise right is not yet a forecast that gets the loss right. On NCA_Kim2011 the two levers agree at cycle 10 to 0.3 mV and the forecast still gets worse, so the difference sits in how each lever extrapolates to cycle 99, about thirty times further in lost lithium than the first ten cycles reach. The practical advice from the previous milestone stands: run the DFN 30 to 50 cycles and fit the law at the physical exponent. The next step tests the extrapolation directly.

Caveats: 40 designs that earlier milestones had already seen, one PyBaMM version, simulated fade only, and inverting the potential on a plateau is ill-conditioned. Code, numbers and the plan: experiments/plateau/ and LOG.md.

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