A plating forecast can borrow how anode layers lag from other designs, on one cell chemistry out of four

A node-gain profile carried from sibling designs cuts OKane2022's ten-cycle plating forecast error from 6.7% to 4.1%, but worsens two other parameter sets.

The cell-fidelity programme forecasts how much lithium a battery design loses to plating by cycle 99 from a ten-cycle simulation. The weak part of that forecast, on the LG M50 parameter set OKane2022, is how fast the anode's local open-circuit potential rises at the layers where lithium plates. The forecast assumes every layer of the anode shifts by the same amount as lithium is lost. The layers don't.

The previous step fitted a correction per layer (a "gain" on the shift) from cycles 3 to 10, and it failed on OKane2022: early on, the layers near the separator lag the shift a lot (gains of 0.1 to 0.8), but that lag shrinks later, so a gain fitted on ten cycles under-reads the cycle-99 rise.

What this step tried

A design search does not start from nothing. Once a few designs of a chemistry have been simulated for 100 cycles, their late per-layer gains are known. So, leaving each design out in turn, we carried the median late gain profile of the other designs in its set into its ten-cycle forecast. The plan, with four hypotheses, was committed before any of it was computed (log, milestone 5zw).

parameter setuniform shiftown ten-cycle gainsborrowed profilemeasured-OCP ceiling
Ai20202.3%2.5%2.2%3.5%
Chen20205.4%3.9%6.5%3.9%
OKane20226.7%6.9%4.1%3.5%
NCA_Kim20113.0%2.3%4.2%1.0%

Set-median error in cycle-99 lithium lost, over 8 to 12 designs per set.

What it found

On OKane2022 the borrowed profile recovers 2.6 of the 3.2 points between the plain forecast and the ceiling you get by measuring the OCP rise outright. The borrowed profile also tracks the OCP rise better than the uniform shift on all four sets. It still makes the forecast worse on Chen2020 and NCA_Kim2011, and it beats the plain forecast on only 23 of 40 designs. On those two sets the separator-side layer's late gain varies a lot from design to design, so a median throws away the part that matters.

One result points somewhere useful. If each design uses its own late gains (which needs cycle 99, so it isn't a forecast), a constant per-layer gain reaches the measured-OCP ceiling on every set. So a constant gain is enough once it's the right one. The open problem is to predict the profile from the design (thickness, porosity, charge rate) rather than averaging it. That's the next step, together with a fresh pool of designs that this programme hasn't seen yet.

A second idea, letting the design's own ten-cycle gains relax toward 1 at a rate chosen on the other designs, also helped OKane2022 (4.7%). It wasn't reliably better than the simpler forecasts on the other three sets, so that hypothesis was refuted.

Limits: all 40 designs were seen in earlier milestones; the pools are small; the per-layer potentials are model internals; one PyBaMM version; modelled fade only. Code and figures: experiments/nodetransfer/.

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