The cell-fidelity programme tries to forecast how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation. The previous step found one fix that works on one parameter set. It takes the profile of how far each anode layer lags the forecast's uniform shift, a "late gain" per layer, from the other fully simulated designs of the same chemistry. On OKane2022 (LG M50) the median profile cut the error from 6.7% to 4.1%. On Chen2020 and NCA_Kim2011 it made things worse. On those sets the gain of the layer next to the separator varies a lot from design to design, and a median throws that variation away.
The obvious next try is to stop averaging over every design and borrow only from the designs that look like the one you're forecasting. Before running anything I wrote down four hypotheses (LOG.md, 5zx). The method takes the three nearest designs in the scaled four-parameter design space (cathode thickness, both porosities, N/P ratio) and uses their median profile. As a second variant, reported but not hypothesised, it fits a per-layer straight line of gain against anode thickness. Each design was held out in turn, on the same 40 PyBaMM DFN designs, with no new simulation.
What came out
Set-median error of the cycle-99 lithium-loss forecast:
| set | uniform lever | set median (5zw) | 3 nearest | thickness line | measured-OCP ceiling |
|---|---|---|---|---|---|
Ai2020 | 2.3% | 2.2% | 2.3% | 3.7% | 3.5% |
Chen2020 | 5.4% | 6.5% | 6.5% | 6.7% | 3.9% |
OKane2022 | 6.7% | 4.1% | 4.5% | 2.7% | 3.5% |
NCA_Kim2011 | 3.0% | 4.2% | 4.1% | 1.5% | 1.0% |
The nearest-neighbour profile failed three of its four pre-registered tests. It beat the set median on one set of four, and it repaired neither of the sets the median hurt. It did keep OKane2022 past the halfway mark to the ceiling. One reason is plain geometry: with 8 to 12 designs in a four-dimensional box, the third-nearest design sits 0.6 to 0.75 of the box's side away, so the "neighbours" aren't near. The neighbours did predict the separator layer's gain about twice as well as the median on Chen2020 and OKane2022. That better gain didn't carry through to the forecast, though. Chen2020's misses come from a few thick designs that are 9 to 17% off whatever lever is used.
The thickness line is the more interesting result. On OKane2022 and NCA_Kim2011 the separator layer's late gain rises with anode thickness, and two parameters per layer, fitted on 7 to 11 other designs, take the forecast to or below the error you get with the measured OCP. On Ai2020 and Chen2020 the gain isn't monotone in thickness (one Ai2020 design has 2.55 where its neighbours have about 1.1), and the line makes things worse. I found this after the plan, so it is a lead, not a result. The next step is to pre-register it on fresh designs, with a check that falls back to the uniform lever when the training designs aren't monotone.
What a design search should do
The advice hasn't changed: fit the lithium-loss law through 30 to 50 cycles. On OKane2022-like cells, a node-gain profile from a few fully simulated designs lets ten cycles do the job.
Limits: all 40 designs were seen in earlier milestones, the pools are small, and several separator-layer gains sit at the ends of the gain grid. The layer OCPs are model internals, everything ran on one PyBaMM version, and the fade is modelled. Code, rows and the figure are in experiments/nodematch/.