Degradation knee

What a knee is (precise)

Take retained capacity on the vertical axis against time or cumulative cycling on the horizontal — the degradation-curve entry covers how those axes are chosen and what conditions them. Most of that line is a gentle near-linear decline. The knee is the transition out of it: the region where fade per year, or per equivalent full cycle, rises well above the mid-life rate and stays there. Naming a specific point requires a detection rule, and several are in common use.

Fit a straight line to the early segment and another to the late segment and call their intersection the knee. Smooth the curve and take the point of maximum curvature. Or set a threshold on the derivative — declare the knee where fade per cycle first exceeds some multiple of the mid-life slope. One dataset, three rules, three different answers, and they need not agree closely. Any statement of the form “the knee is at N cycles” is incomplete until the rule that produced N is stated alongside it.

The maximum-curvature family carries an extra dependency worth understanding, because it is a geometric property rather than a physical one: curvature is not invariant when the axes are rescaled. Stretch the vertical axis, or compress twenty years into the same plot width, and a gentle bend acquires a sharp corner without a single data point moving.

Change the horizontal axis from calendar years to equivalent full cycles and the same history re-shapes itself according to how the cycling was distributed over the years — a plant that cycled harder late in life pushes its own bend leftward or rightward depending on which currency it is measured in.

Two-line-intersection and derivative-threshold rules are less sensitive to aspect ratio, but they trade that for a segmentation choice: where the early segment is judged to end determines what the late segment is fitted to. Knee location is therefore a joint property of the cell, the plot and the rule, and only the first of those three belongs to the battery.

Three distinctions keep the term clean. A knee is not end of life, which is a contractual threshold written against a retention percentage and a duty. A knee is not sudden failure — a cell past its knee still charges and discharges, it simply loses capacity faster per unit of use.

And a knee in capacity is not the same event as the bend in the resistance trajectory, sometimes called an elbow, which follows its own timeline: resistance can turn upward before, with, or after capacity steepens, and a design that fades gracefully in energy can still lose power capability first. Three curves, three shapes, one asset.

Candidate mechanisms, none of them promoted to a cause

Every mechanism on the candidate list is already running during the linear middle; what changes is the regime, not the cast. Lithium plating is the most cited: metallic lithium deposited on the anode during cold or fast charging, or near full charge, removes cyclable inventory directly, and the rough deposits present fresh surface for further film growth.

Electrolyte depletion is the second: film growth consumes electrolyte over the whole life, and once the remaining volume no longer keeps the porous electrodes properly wetted, ionic transport degrades across the cell rather than merely costing a little more inventory. The third group is mechanical and electrode-level — particle cracking under repeated expansion and contraction, active material losing electrical contact with the conductive network, and reaction products filling the pore space that transport depends on. None of these is exotic and none of them switches on at a particular date.

What makes any of them knee-shaped is that they stop behaving like steady taxes. Two shapes of explanation do that work. The first is positive feedback, and it is the framing our degradation-mechanisms article uses: plating roughens the anode, roughness grows more film, thicker film raises resistance, resistance makes local heat, heat accelerates everything, and the loop closes.

The second is reservoir exhaustion: a finite quantity — electrolyte volume, accessible porosity, spare lithium inventory, anode capacity balance against the cathode — is drawn down smoothly for years, and the curve bends when what remains stops being sufficient for the duty. Both reproduce the observed shape, both are consistent with the same benign linear middle, and they are not mutually exclusive: a depletion threshold is a convenient way to start a feedback loop.

The honest position is that which mechanism leads is a function of cell design and duty rather than of chemistry alone. Electrode loading, porosity, electrolyte fill volume and the balance between anode and cathode capacity all move the thresholds; cold or fast charging biases toward plating, sustained heat and high resting state of charge bias toward electrolyte consumption. A knee observed in one cell format under one accelerated protocol identifies the mechanism for that cell under that protocol, and transfers to another product only as a hypothesis.

The deeper limit is evidential: the shape of a capacity curve does not identify its own cause. Distinguishing plating from depletion from mechanical loss takes differential-voltage or incremental-capacity analysis, resistance trending, and eventually teardown of aged cells — not a steeper line on a plot. The practical rule for a project is to name the candidates, resist adopting one as the cause, and require the distinguishing evidence before a model is allowed to lean on a particular mechanism.

Why models place it differently

Degradation models come in two broad families, and the knee is where they part company. Empirical fits express fade as a function of time, throughput, temperature and depth of discharge, with coefficients regressed against cycle-test data. Such a model can only produce a knee if its functional form permits one — a linear or square-root-of-throughput term cannot bend upward however the coefficients are chosen — and only if the data it was fitted to contains one.

Physics-based and semi-empirical models can generate a knee from their internal states, tracking plating overpotential or consumed electrolyte, but the parameters that decide when those states cross their thresholds are themselves fitted to the same finite body of test data. Neither family knows more about the far end of life than the experiments behind it, and both must extrapolate to reach it.

The data behind them is bounded by the clock. Cycle-life testing takes real years, so cells are aged at elevated temperature or higher C-rate to compress the schedule, and the compression is the problem. Accelerating a test changes the relative pace of the mechanisms, and can change which one crosses its threshold first — an elevated-temperature run may reach electrolyte exhaustion long before a field asset ever would, while a field asset charging cold in winter has a plating exposure the hot lab cells never see.

The rule of thumb that calendar fade rates roughly double for every 10 °C of sustained cell temperature describes a rate, and cannot be used to slide a knee location from lab conditions to site conditions as if the two curves were the same shape stretched.

The consequence is a spread between credible models, and the spread is real rather than a sign that one of them is wrong. Two vendors can publish trajectories agreeing to within a percent across the warranted region — the region where both have measurements — and disagree by years about where fade steepens, because past that region each is showing the tail of its own functional form.

Handle it the way the uncertainty deserves: carry the knee as a sensitivity case rather than a line item, run the augmentation schedule and any life-extension or residual-value assumption against an early-bend scenario as well as the base case, and treat a model that reports a precise knee date without an interval as having hidden its extrapolation rather than resolved it.

Why the warranty stops short of it

The capacity warranty is a year-by-year retention table running down to the contractual floor — commonly 65–70% of beginning-of-life capacity — conditioned on an envelope of cycles or throughput per year, depth of discharge, temperature and state-of-charge dwell. Both terminations matter here.

The retention table ends at the floor, and the floor is placed where the supplier expects the plant to still be on the predictable stretch of the curve; the throughput cap ends the guarantee at a cumulative usage the test programme can stand behind. The knee, by construction, is expected to sit past one or both. So the document that appears to warrant the shape of the degradation curve is silent about the knee: no row of the table locates it, and no remedy attaches to reaching it.

That allocation is defensible rather than evasive — a supplier prices what it can evidence, and it cannot evidence the region past its own test data. But it does place the risk somewhere specific.

It falls on whoever owns the asset in the years past the floor or past the envelope: the operator who re-optimises mid-life into heavier cycling than the warranted duty, the owner weighing life extension against augmentation or repowering, and anyone screening aged cells for a second-life application, where the buyer is deliberately purchasing hardware in the part of the curve the original warranty declined to describe.

Two practical consequences follow for documents and operations. First, the envelope clauses are the fence, and crossing them does more than forfeit a remedy — heavier duty, hotter cells or more time spent full re-derives the whole trajectory including where it steepens, so the exposure and the warranty lapse arrive together rather than one after the other.

Second, since the retention table offers no early warning, the only warning available is the plant's own telemetry: the resistance trend at fixed temperature and state of charge, the divergence between best and worst cells, and the sequence of capacity-test points read as a slope rather than as pass or fail against a floor. A retention series that keeps clearing the warranted line while its slope steepens test over test is the signal worth escalating, and it is visible years before any threshold is crossed.

Reading a curve that shows none

Most published degradation exhibits show no knee, and the reason is usually structural rather than reassuring. A plotted curve can only display a bend inside its own horizontal range, and a table that ends at the end-of-life year or the warranted cycle cap ends before the region where a bend is expected. A fitted line whose functional form is monotone-linear or square-root cannot render one at any range.

And a linear extrapolation of data taken before a knee projects no knee by construction — the absence is an output of the method, entered as an assumption. “Straight to year 20” is therefore a statement about the plot's range, its functional form, the duty it was evaluated at and how long the cells actually ran, and it becomes a statement about the product only when those four are known.

Five questions convert the exhibit into evidence, and they are all about the tail rather than about the curve as a whole. How long did the longest-running cells in the underlying programme actually run, in calendar time and in cycles, and to what retention did they get? Where on the plotted line does measurement stop and extrapolation begin — is that boundary marked?

At what temperature and C-rate were those longest tests run, relative to the site's design conditions? Did any tested cell bend, and how were those cells treated in the fit rather than set aside as outliers? And what rule would the supplier itself apply to declare a knee in field data, since a threshold nobody has defined cannot be argued about later. A supplier who can answer these has a test programme; one who answers by re-showing the curve has a chart.

The error runs in the other direction too: an apparent knee in operating data is often not one. Before concluding that the chemistry has turned, rule out the usual impostors. A change to the usable state-of-charge window remaps how much of the pack the capacity test can reach. A revised test protocol, a different rest period before the test, or an uncorrected temperature difference between this year's test and last year's moves the measured number without moving the cells. Augmentation energy entering the retention arithmetic mid-life redefines what the percentage is a percentage of.

And divergence produces a genuine bend at the wrong level entirely — under weakest-cell limitation, the most degraded cells clamp the string's charge and discharge limits, so pack usable energy can steepen while average cell capacity carries on down its straight line. A pack-level knee sitting above linear cell-level fade is a balancing and maintenance finding, and telling the two apart takes cell-level data, which is exactly the data that stops being collected once a fleet settles into routine operation.