The cell-fidelity programme forecasts a lithium-ion cell's 100-cycle lithium-plating loss from its first ten cycles, so an electrode design search does not have to simulate all 100. The test bed is PyBaMM's full DFN model with SEI growth and lithium plating, on four parameter sets. The previous milestone got LG M50 (Chen2020, OKane2022) to within about 10%. It did that by reading both levers of the plating law at the charge window the cell actually reaches. That window is set by one effective overpotential, η, measured from the first ten cycles' charge capacity. The same change made Ai2020 worse, from 2.5% to 7.6%. The explanation offered then was that Ai2020's η drifts over the 100 cycles while the law holds it constant.
Three ways to fix it, none of which worked
Milestone 5ze tested that explanation directly. It first tried three fixes on the 384 designs earlier milestones had already simulated:
- Follow η's ten-cycle trend. The trend over-extrapolates. On LG M50 it puts the cycle-99 η 20 to 45 mV too low, and the LG M50 error rises from 6 to 10% to 15 to 18%.
- Choose the reading per design. Pick the open-circuit or the η-window reading by which one predicts cycle 10 better from cycles 2 to 7. On
Ai2020andOKane2022this picks the wrong one for most designs. The two readings only part after cycle 10. - Split η in two. A symmetric η was an assumption. The mean discharge voltage (the discharge energy over the discharge capacity) fixes where the window sits as well as how long it is. So each design gets a discharge-side and a charge-side overpotential from data it already has. The charge side came out within a few millivolts of the symmetric value, and the forecast barely moved.
The fourth check was the telling one. It is an oracle, not a forecast: give the law the η the cell really needs at cycle 99. If drift were the problem, this would fix it. It made the forecast worse on every seen pool.
The fresh test
The milestone did not propose a new law. It pre-registered five claims about these findings and ran 128 new designs to check them: a new random seed, and currents no earlier milestone had used. All five held.
| pool | open circuit (5zc) | η window (5zd) | split η (5ze) | true drift (oracle) |
|---|---|---|---|---|
Ai2020@0.6125x | 2.4% | 6.3% | 5.9% | 11.1% |
NCA_Kim2011@0.5625x | 1.3% | 2.3% | 2.3% | 4.0% |
Chen2020@1.28125x | 19.1% | 10.9% | 10.2% | 15.5% |
OKane2022@1.3125x | 16.0% | 9.5% | 9.3% | 12.6% |
The figures are the median absolute log error of the predicted cycle-99 lithium loss, 32 designs per pool. The LG M50 result from 5zd replicates. The split overpotential's charge side is within 1 to 5 mV of the symmetric one, and the discharge side is about a third lower. The oracle loses on all four pools.
What it means
The overpotential is measured correctly. Measuring it better, or even knowing its future, does not help, so the remaining error is in the structure of the law. On LG M50, a constant η appears to make up for something the law leaves out. The leading candidate is the plating kinetic exponent, which the law takes from one parameter set and applies to all four. The next milestone estimates it per design from the same ten cycles.
The same caveats apply as before: one current per set, one PyBaMM version (26.10.1.1), and modelled fade, not measured cells. The code, the 128-run bench and the pre-registration are in the project directory and its research log.