A battery's charge-capacity fade follows its electrode balance, but that doesn't fix the plating forecast

The electrode-balance window forecasts cycle-99 charge capacity within 4 to 8% from ten cycles, but the plating forecast's error stays: it is the lever.

The cell-fidelity programme tries to forecast how much lithium a battery design loses to plating by cycle 99, using only a ten-cycle simulation. The last step fitted straight lines through cycles 2 to 10 to two of the forecast law's inputs and found the lines don't hold. The worse of the two was the capacity factor: the charge the cell takes at cycle 2 divided by what it takes at cycle n. Its line under-read cycle 99 by 16 to 51%, because the factor curves upward as lithium is lost.

That curve has a physical reason that is easy to compute. When a cell loses cyclable lithium, its working window slides along the two electrodes' open-circuit potential curves. Those curves are not straight, so the window narrows faster the further it slides. Milestone 5r of this programme already had that map, from the design and the two curves alone. This step tested it as the capacity-factor forecast.

What I did

For the same 40 PyBaMM DFN designs (four parameter sets: Ai2020, Chen2020, OKane2022 and NCA_Kim2011), I took each cycle's measured charge from cycles 2 to 10 and found the overpotential η that shrinks the open-circuit window to that charge. I fitted a line to η against lithium lost, then asked the window what charge it would hold at each later amount of lithium lost. Nothing past cycle 10 went in. The hypotheses and their bars were committed before anything was computed.

The capacity factor: yes

At cycle 99 the window forecast is within 3.6 to 7.6% of the measured factor on all four sets. The straight line missed by 16 to 51%.

setstraight linewindow, η linewindow, η heldmeasured
Ai20201.201.862.071.92
Chen20201.351.411.691.54
OKane20221.341.401.641.53
NCA_Kim20111.591.872.002.00

The column I didn't expect to matter is "η held". That is the window with the overpotential fixed at its ten-cycle mean, and it is the forecast law's own capacity factor. It is within 2.4 to 7.1% of the measured value. So the law already had the right shape, and the previous step's failure was the straight line, not the law.

There was a second surprise. I expected resistance growth to push η up and end each charge earlier. It does the opposite here. η falls by 6 to 18 mV over the first ten cycles on every set and then levels off, and that is why the η line ends up low by cycle 99.

The forecast: no

Fed to the law together with the node lever from milestone 5zn, the window factor beats the law's own lever on only 14 of 40 designs. Set-median cycle-99 errors were 2.8% on Ai2020 (law's lever 8.3%), 5.4% on Chen2020 (8.2%), 8.1% on OKane2022 (5.9%) and 5.7% on NCA_Kim2011 (2.2%). That is within half a point to a point of the node lever with the law's own factor. Getting the capacity factor right changed almost nothing.

So on the LG M50 sets, the error the ten-cycle forecast still has comes from the lever, the law's driving force for plating, and not from the capacity factor. The next step is to forecast the plating overpotential's path from the cell's internal states at cycle 2, not only the anode's open-circuit potential.

What this does and doesn't show

The 40 designs were all seen in earlier steps, so this checks a known pool and isn't a held-out test. η is a symmetric shift of both voltage cut-offs, which stands in for the real charge overpotential. Everything is modelled fade in one PyBaMM version. For anyone running a design search with this law, the practical advice hasn't changed: fit its line through 30 to 50 cycles, where the error is 1.1 to 3.2%.

Code, results and the full log: research/cell-fidelity, experiments/window/.

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