Same designs, opposite verdict: the plating exponent is read through the electrode after all

On 40 PyBaMM DFN designs the plating-weighted OCP rise sets the plating exponent within 3 to 12% on three of four cells, reversing a cross-pool comparison.

The cell-fidelity programme forecasts a battery design's 100-cycle lithium loss from its first ten cycles. The forecast needs one kinetic exponent per cell type, and the open question for several milestones has been where that exponent comes from.

The last milestone found that plating in PyBaMM's DFN model is spread through the negative electrode, not piled up at the separator. It then read the potential lever where the plating is and concluded it did not explain the exponent. That conclusion rested on a comparison across pools: the depth ratio came from 40 designs, and the exponent it was checked against came from 768 other designs run at other currents.

This milestone redid the comparison on the same 40 designs. The x-resolved records already held everything the forecast law needs (lost lithium per cycle and the constant-current charge time), so no new simulation was run. The hypotheses were committed before any number was computed.

What came out

cellexponent neededseparator-node readingplating-weighted reading
Ai202018.843.625.2
Chen2020 (LG M50)36.640.537.7
OKane2022 (LG M50)32.337.534.2
NCA_Kim201123.617.326.4

Set medians, per volt. The reading is αF/RT = 25.3 V−1 times the ratio of the measured end-of-charge OCP rise to the law's own lever rise.

  • The plating-weighted reading is closer than the separator node's on 37 of 40 designs (one-sided sign test p = 10−8).
  • Within a cell it ranks designs: Spearman 0.83 to 0.95 on three of the four sets. On NCA_Kim2011, where the needed exponent barely varies, it does not.
  • Used as the exponent, it cuts the 100-cycle forecast error on LG M50 by about three times (0.082 to 0.025 and 0.059 to 0.019 in |log ratio|).
  • Ai2020 stays about a third high, mostly from two of its eight designs.

What it means, and what it does not

The exponent a plating forecast needs is not a free calibration constant: on three of four cells it is the physical one, scaled by how much faster the potential rises where the plating happens than the forecast's uniform-stoichiometry lever assumes. The last milestone's reading pointed the other way only because both sides of its ratio came from different currents.

It is still a diagnosis. The lever uses cycle 99's potentials, so it cannot yet forecast anything from ten cycles. That is the next step. Three of the pre-registered hypotheses came out inconclusive, including the two that checked whether the earlier pools transfer to these designs: they do not on Ai2020 and Chen2020. All numbers are from PyBaMM 26.10.1.1 with one set of plating parameters on every cell, so this is modelled fade, not measured cells.

Details: results, log entry.

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