A battery rarely warns you before it gets old. Capacity slips away quietly, cycle after cycle, until one day the decline turns sharp. The promise of predictive AI is to see that future early — to read the faint signals in a cell's first weeks of life and forecast how it will age long before any owner would notice.
How a lithium-ion cell actually ages
Ageing is not one process but several, running in parallel and feeding one another. The most universal is the slow growth of the solid-electrolyte interphase — the SEI, a thin protective film on the anode. The SEI is necessary; without it the electrolyte would react uncontrollably. But it never stops growing entirely, and every time it thickens it consumes a little cyclable lithium. That lithium is the cell's working currency, so as the SEI grows, capacity quietly shrinks.
Other mechanisms pile on. Under fast charging or cold temperatures, lithium can plate as metal on the anode instead of inserting into it — sometimes reversibly, sometimes not, and occasionally in ways that threaten safety. Active material can crack, detach, or lose electrical contact, a process broadly called loss of active material. The electrolyte itself is depleted by side reactions, drying out the cell's internal pathways. And as all of this accumulates, internal resistance climbs — an impedance rise that makes the battery feel weaker even when capacity numbers look acceptable.
Why fade is so hard to see early
The frustrating part is timing. For most of a battery's life, degradation is gradual and nearly linear — a percent here, a percent there. Then, often without obvious warning, the curve bends. Researchers call this the degradation knee: a point where one mechanism, frequently lithium plating or a collapse of usable lithium inventory, begins to accelerate the others. After the knee, a cell can lose more capacity in a few months than it did in years.
By the time a battery looks old, it has already been ageing for a long time — the signal was there all along, just too quiet to read by eye.
This is why simply watching capacity is a poor early-warning system. The number you can measure most easily is the one that stays reassuring the longest.
From fade curves to physics-informed AI
Modelling ageing has progressed through distinct generations, each correcting a weakness of the last.
The empirical and circuit eras
The earliest approach was the empirical fade curve — fit a simple mathematical shape (often a square-root-of-time or power law) to observed capacity loss and extrapolate. It is easy and sometimes useful, but it knows nothing of why a cell ages, so it fails the moment conditions change. Equivalent-circuit models added a layer of physical intuition, representing the cell as resistors and capacitors whose values shift as it degrades. They capture impedance behaviour well but still treat the chemistry as a black box.
Physics and machine learning together
Physics-based electrochemical models go deeper, simulating lithium transport, reaction kinetics and SEI growth from first principles. They are accurate and explanatory — but slow to compute and demanding to calibrate. The current frontier, physics-informed machine learning, blends the two: data-driven models learn from large volumes of real cycling data while being constrained by the known physics of degradation. The physics keeps predictions honest; the learning lets them adapt to messy, real-world behaviour.
The signal hides in the early cycles
The most striking finding from the battery research literature is that a cell's long-term fate is partly written in its first weeks. Features extracted from early-cycle data carry a strong predictive signal — often enough to forecast end-of-life well before capacity has visibly dropped. The richest signals include:
- Discharge voltage curves — subtle changes in the shape of the voltage-versus-capacity curve between early cycles reveal degradation that the capacity total alone hides.
- Differential capacity (dQ/dV) — the position and height of peaks in this curve map directly to phase transitions in the electrodes, so their drift exposes loss of lithium and active material.
- Voltage relaxation — how a cell's voltage settles after a rest period reflects its internal kinetics and is sensitive to early plating and impedance change.
- Impedance measurements — resistance growth tracked across frequencies separates film growth from contact loss, distinguishing degradation modes that look identical on a capacity plot.
None of these requires waiting for the knee. They are diagnostic fingerprints of mechanisms already in motion, and a well-trained model can read them long before a user senses anything wrong.
Why the genome makes the prediction generalise
A model fitted to one chemistry, one cell design and one duty cycle can be impressively accurate — within that narrow world. Move it to a different cathode, a colder climate or a faster charging profile, and its forecasts drift. Ageing depends on what the cell is made of, and a model blind to composition cannot transfer what it learned.
This is where grounding predictions in a cell's measurable genome matters. When a model is connected to the actual material fingerprint — structure, composition, purity, electrochemical signature — it is no longer guessing across chemistries. It is reasoning about a specific cell with a known make-up. Predictions built on the genomic record generalise far better, because they explain ageing in terms of cause rather than correlation.
What changes when you can see ageing coming
Accurate early forecasts change decisions across the battery's life. Warranties can be priced on evidence instead of conservative worst-case padding. A used pack can be valued for second-life service on its documented trajectory rather than a rough guess, unlocking storage value that is otherwise written off. Fleets can move from fixed replacement intervals to predictive maintenance — servicing or retiring a battery when its own data says it is near the knee, not when a calendar says so.
The shift is from reacting to ageing to anticipating it. A battery that can tell you, early and credibly, how it will grow old is a battery you can plan around — and that foresight, more than any single chemistry, is what turns a fleet of black boxes into infrastructure you can trust.