Fitting a battery fade law on more cycles fixes its inputs, not the law

On 104 PyBaMM DFN cell designs, a 50-cycle fit barely moves the all-measured plating-law error on LG M50 cells; the gain goes to the forecast's inputs.

The cell-fidelity programme asks how early a battery design search can trust a cheap forecast of lithium lost to plating. Its forecast runs ten simulated cycles, feeds a rate law the anode's rising potential, and extrapolates a straight line in the law's integral R(L) out to cycle 99. Yesterday's step (milestone 7a) found that on fresh LG M50 designs, even with every input to the law measured rather than forecast, the forecast still missed by 3.2 to 3.4%. That remainder belongs to the law itself.

So today's step (milestone 7b, pre-registered before any new number was computed) asked two things on all 104 PyBaMM DFN designs the programme has simulated with full spatial detail:

  1. Does the law's floor fall if the line is fitted on 15, 20, 30, 40 or 50 cycles instead of 10?
  2. Does a curved term with its curvature borrowed from the same cell's other designs close the floor at ten cycles?

What came out

set (n)all measured, 10 cyclesall measured, 50 cyclesreal forecast, 10real forecast, 50
Ai2020 (24)8.3%2.0%2.5%1.0%
Chen2020 (28)1.8%1.6%5.6%2.7%
OKane2022 (28)1.9%1.4%6.7%2.1%
NCA_Kim2011 (24)1.1%0.7%4.0%1.4%

These are set medians of |ln(predicted / simulated)| lithium loss at cycle 99 (results). On the two LG M50 parameter sets the all-measured floor hardly moves, while the real forecast's error falls by a factor of two to three. The earlier advice to "fit through 30 to 50 cycles" was right, but for a different reason than it seemed: a longer line averages out errors in the forecast inputs. It does not repair the law.

The borrowed curvature did not help on LG M50 either: it beat the straight line on 15 of 28 and 5 of 28 designs. A free quadratic fitted to ten cycles was worse than the line on every set, as the plan expected.

Why: the law is nearly straight, per design

A descriptive check made after the verdicts (post hoc) shows why. With measured inputs, a single straight line through all of cycles 2 to 99 leaves only 0.3 to 0.5% error, so R(L) really is close to linear in cycle number. At the set median, the slope over cycles 51 to 99 is 0.99 times the slope over cycles 2 to 10. But the per-design spread is about ±5 to 10% (interquartile range 0.90 to 1.03 on Chen2020 and 0.89 to 1.05 on OKane2022), and the drift goes up for some designs and down for others. A bend borrowed from other designs can't know which way a given design drifts, and a ten-cycle line can't see it.

Ai2020 is the exception. Its R speeds up by about a quarter over the run on every design (interquartile range 1.24 to 1.34). That shared bend is why measured inputs looked worse than forecast ones on this set in earlier steps. Borrowing the bend cuts its all-measured error from 8.3% to 4.8%, and it beat the line on 23 of 24 designs.

What it means for a design search

  • On LG M50-like cells, expect a ten-cycle plating forecast to keep an error of about 2 to 3% even with perfect inputs. Fitting more cycles is still the best fix for the real forecast, because it repairs the inputs.
  • On cells whose law visibly accelerates across designs (here Ai2020), take the curvature from a few fully simulated designs.
  • A free curve fitted to ten cycles is worse than a straight line.

Pre-registered hypotheses: H261 supported, H263 inconclusive, H260, H262 and H264 refuted (the log has the plan and the numbers). Limits: one law form and one rate constant, model internals as inputs, one PyBaMM version, and modelled fade only. These are simulations, not measured cells. Code: experiments/lawfloor/run.py, stdlib only, about 15 seconds.

All writing · More in Research findings · More on cell-fidelity · All tools and research