What state of health actually means
State of health is a ratio, not a measurement. Nothing in a battery reports its own health, so every SoH figure is the output of an estimator that someone designed, tuned and calibrated.
Two definitions dominate, and they don't move together:
Capacity-based SoH compares present usable capacity to rated capacity at beginning of life. This is the number that shows up in warranties and second-life valuations.
Resistance or power-based SoH tracks internal resistance growth and remaining power capability. A cell can hold most of its capacity while losing the ability to deliver rated power, which matters enormously for frequency response and fast charging, and not much at all for a four-hour arbitrage asset.
Training that teaches only the capacity definition leaves people unable to explain why two SoH figures for the same asset disagree, which is one of the most common arguments in a commercial review.
Why the topic can no longer be treated as advanced
Under Regulation (EU) 2023/1542, since 18 August 2024, battery management systems in stationary battery energy storage systems, LMT batteries and EV batteries must hold up-to-date data on the parameters that determine state of health and expected lifetime, set out in Annex VII, with read-only access for the legal purchaser of the battery. Those parameters go well past capacity: remaining power capability, round trip energy efficiency, self-discharge behaviour and internal resistance or impedance all appear.
That turns SoH from an engineering nicety into a compliance artefact and a contractual one. Anyone writing or reviewing a supply agreement for the EU market now needs to understand what the number means and how it was produced.
The degradation mechanisms a curriculum has to name
Capacity fade and resistance rise are symptoms. Training that never names the underlying mechanisms leaves engineers guessing at causes:
Solid electrolyte interphase growth consumes cyclable lithium at the anode and drives most calendar ageing. Under diffusion-limited conditions it follows roughly square-root-of-time behaviour, which is why storage duration and storage temperature matter as much as cycling for some assets.
Lithium plating occurs at low temperature, high charge rate, or high state of charge, and is largely irreversible once it forms dead lithium. It's the mechanism behind most fast-charge damage arguments.
Loss of active material through particle cracking, binder failure and contact loss removes host sites for lithium.
Loss of lithium inventory and loss of active material are the two lenses that let an engineer interpret a diagnostic curve rather than just read a number.
Electrolyte decomposition, gas generation, transition metal dissolution and current collector corrosion round out the set.
Then there's the knee: many cells age gently and then accelerate, often driven by plating and pore clogging in combination. Any degradation model built on the early linear region alone will overstate remaining life, and any training that presents ageing as a straight line teaches a costly mistake.
Three families of degradation model
| Approach | How it works | Good for | Where it breaks |
|---|---|---|---|
| Empirical and semi-empirical | Fits capacity loss to stress factors: temperature via Arrhenius terms, depth of discharge, C-rate, mean state of charge, time and throughput | Warranty modelling, project finance, quick fleet estimates | Extrapolation beyond the tested envelope, and knee onset |
| Physics-based | Electrochemical models (single particle, pseudo-two-dimensional) with side reactions such as SEI growth and plating | Design decisions, root cause work, new chemistry evaluation | Parameterisation cost, compute load, needs cell teardown data |
| Data-driven | Features extracted from partial charge curves, resistance trends or relaxation voltage, fed to statistical or machine learning models | Fleets with dense telemetry, early-life life prediction | Generalising to chemistries, formats or duty cycles outside the training data |
Most production systems blend them: a physics-informed model constrained by empirical stress factors, corrected by field data. Training that presents only one family produces engineers who can't read a vendor's methodology.
Diagnostics that separate causes from symptoms
Incremental capacity analysis (dQ/dV) and differential voltage analysis (dV/dQ) turn a slow charge curve into peaks that shift and shrink in characteristic ways. Reading them lets you distinguish loss of lithium inventory from loss of active material, which changes what you do next. Electrochemical impedance spectroscopy and pulse-resistance tests fill in the resistance picture.
These are teachable in an afternoon and rarely taught at all outside a lab context.
Why BMS calibration is the topic everyone skips
State of charge estimation in most systems starts with coulomb counting, integrating current over time. Current sensors have offset and gain error, so the estimate drifts, and the drift compounds. Everything else in the stack exists to anchor it.
Open circuit voltage lookup after a rest period is the usual anchor, since OCV maps to state of charge for a given chemistry and temperature. Model-based observers, most often an extended or unscented Kalman filter over an equivalent circuit model, blend the two continuously.
Two practical problems dominate the field:
LFP has a very flat OCV curve across the middle of its range, so voltage tells you almost nothing about state of charge between roughly 20% and 80%, and voltage hysteresis makes it worse. Systems relying on OCV correction need to reach a charge or discharge extreme to recalibrate, which is why LFP BESS operating strategies often schedule periodic full charges.
Capacity re-estimation needs a deep, slow, controlled cycle, and grid-connected assets are rarely idle enough to give you one. So capacity estimates get inferred from partial cycles, with the accuracy penalty that implies.
Add firmware updates and BMS resets that clear learned parameters, temperature-dependent parameter drift, and cell-to-cell spread that means pack SoH is not the average of its cells, and you have a set of failure modes that shows up in real disputes. Training that never mentions them leaves people unable to interrogate an SoH report.
What "battery training algorithms" should mean in a curriculum
The phrase covers two things that are easy to confuse. One is the estimation algorithms running inside a BMS: coulomb counting, OCV correction, Kalman filtering, equivalent circuit and reduced-order electrochemical models. The other is training machine learning models on cycling data to predict remaining useful life. A current programme should cover both and be explicit about which one it's discussing at any moment, because the vocabulary overlaps and the engineering does not.
How BatteryMBA covers it
BatteryMBA is a 12-week CPD-accredited live online programme run by Battery Associates. Degradation, cell chemistry and battery management sit inside a wider value chain view, so participants see how an SoH estimate becomes a warranty clause, a second-life valuation and a regulatory obligation rather than treating it as an isolated modelling exercise.
Lectures are taught by practitioners currently working in the industry, across raw materials, manufacturing, integration, BESS, EVs, recycling, policy and investment. Weekly office hours run across multiple time zones, and every session is recorded. Expect 2 to 3 hours a week, or 4 to 5 with the optional case study track.
Participants also get access to Battery101, an on-demand primer covering lithium-ion fundamentals, so a mixed-background cohort starts from the same baseline.
C18 runs 14 September to 5 December 2026. Tuition is €2,900.
Enrol in C18 · See past lecture topics
Informational and educational content only. Not professional, financial, legal, or engineering advice.