The capacity knee of a physics-based Li-ion model is governed by one unmeasured exponent.

<title>Abstract</title> <p>Physics-based degradation models are valued because their parameters correspond to physical quantities. For the capacity “knee” — the late-life acceleration that sets end of life — we show that this promise fails in a specific, checkable way. Using PyBaMM with the O’Kane stack on an LFP/graphite cell, we sweep eight parameters over 512 volume-closed simulations. Knee severity tracks a single quantity: m_LAM, the exponent of the stress-driven loss-of-active-material rate law (Spearman rho = + 0.921, p &lt; 1e-16, n = 114). The conventional rate constants — SEI growth, lithium plating, particle cracking — change how much a cell fades but not the shape of its fade (all |rho| ≤ 0.144, p ≥ 0.126). A 1024-point volume-closed control holding m_LAM at its assumed value of 2.0 reaches a maximum knee ratio of 1.76 across 982 successful runs, against an observed median of 7.23 in 131 MIT/Severson cells. m_LAM is recorded as “Assumed” in its source parameterisation, with no measured value known to us for graphite or LFP. Inverting knee(m_LAM) against each of 35 exact fast-charge policies gives a required exponent of 13.35-18.00 (median 14.31), at least 6.7 times the assumed value in every policy. m_LAM is further confounded with the negative-to-positive capacity ratio: parameter sets differing by a factor of 1.5 in m_LAM reproduce the observed knee to within 25 %. The model can therefore be fitted to an observed knee but cannot predict one, and knee agreement is not evidence of predictive capability unless m_LAM and N/P are reported together.</p>