The cell-fidelity programme asks how much of a lithium-ion cell's 100-cycle lithium-plating fade can be forecast from its first ten cycles, so a design search does not have to simulate all 100. The model is PyBaMM's full DFN with SEI growth and lithium plating. The hard case is LG M50 (Chen2020 and OKane2022 parameters), where every law so far has over-predicted the loss.
The two levers
The forecast law comes from the plating kinetics. Plating per cycle grows with how long the charge lasts and shrinks as lost lithium raises the negative electrode's potential at the end of the charge. The previous milestone, 5zc, fixed the first lever. It measured an effective overpotential, η, from each design's first ten cycles of charge capacity: the charge stops when the open-circuit voltage is η short of the cut-off. It still read the potential lever at the open-circuit cut-off.
That is inconsistent. If the charge stops η short, the electrode's open-circuit state at the end of the charge is the top of that shorter window, so milestone 5zd reads the potential rise there too. It is the same η, so no new number is fitted.
What the fresh pools said
The plan was committed before any run: 128 new electrode designs at currents no milestone had used, simulated to cycle 100 (about 20 minutes on four cores).
| pool | 5zb | 5zc | 5zd |
|---|---|---|---|
Chen2020 (LG M50) | 0.270 | 0.202 | 0.104 |
OKane2022 (LG M50) | 0.259 | 0.167 | 0.064 |
NCA_Kim2011 | 0.028 | 0.033 | 0.020 |
Ai2020 | 0.040 | 0.025 | 0.076 |
The figures are median |ln(predicted / simulated)| lithium loss at cycle 99, so 0.10 is about 10%. On LG M50 the new reading wins on 59 of 64 designs. It also forecasts the measured surface-potential rise next to the separator better than the open-circuit reading did (17 to 21 mV of error against 27 to 35 mV).
The cost falls on Ai2020. Its error triples to 7.6%, 0.051 worse, just past the 0.05 harm limit the plan set. Its overpotential drifts by about 20 mV over 100 cycles, so a fixed η overstates the late potential rise. Rank correlation on LG M50 also dips (0.97 and 0.94), so the change fixes the size of the loss more than the order of designs. Four of five hypotheses were supported and one was mixed.
The next step is to let η follow its own ten-cycle trend and see whether the Ai2020 miss closes without losing the LG M50 gain.
Limits: modelled fade in one PyBaMM version; one current per set; the form was chosen after looking at 256 earlier designs and seven candidate forms, all listed in the dev table.
Code and table: experiments/consistent/. Research log: LOG.md.