Thermal runaway

Between the moment a lithium-ion cell begins to fail and the moment it burns there are four distinguishable stages. Each releases a different gas picture. Measure one gas and you see one stage — usually the third, by which point the cell has already opened. Measure the pattern and you see the second.

The four stages and their gas signatures

The following follows the work of Essl and colleagues, published in Batteries in 2021, and the NANOZ analysis for automotive batteries derived from it.

The temperature jump is striking. Roughly a hundred degrees separate stage b from stage c, and with them — depending on the failure mode — minutes to hours. Between stage c and stage d lies the difference between an opened cell casing and a fire.

Why a hydrogen sensor arrives too late

The obvious approach is to measure hydrogen. It appears at stage a already through unwanted electrolysis, it is easy to detect, and sensors for it have existed for decades.

The problem is not detection but discrimination. At stage a hydrogen forms at cell level in small quantities — where it has not yet travelled to a sensor on the pack enclosure. By the time enough hydrogen is in the measurement volume to cross a defensible threshold, the cell has as a rule already reached venting. Thermal-conductivity devices also measure an aggregate: they register that the thermal conductivity of the gas mixture has changed, not what changed it.

That produces a dilemma known in practice as errors of the first and second kind. Set the threshold low and every external interferent triggers an alarm — a vehicle parked next to a running combustion engine in an underground garage, or a cloud of cleaning agent in the workshop. Set it high and you detect the event only once it is visible anyway. Neither is usable for a battery management system.

What actually happens at stage b

At stage b the electrolyte of damaged cells vaporises. That is the earliest point at which cell damage becomes noticeable outside the cell — and simultaneously the hardest to detect, because no single marker gas escapes but a mixture.

Worse: the composition depends on the cell type. Four common electrolyte formulations produce four different gas pictures.

In every case water, carbon dioxide, carbon monoxide, ethane, hydrogen and ethylene are added to these. A sensor calibrated on a single marker gas would therefore have to be re-specified for every cell type — and would be blind to a cell type the manufacturer switches to later.

This is precisely why the task demands an electronic nose rather than a single-gas sensor: the information is not the concentration of one substance but the ratio of several to one another. That ratio is characteristic of “electrolyte is vaporising” and differs from “a diesel is driving past outside”.

Why MOx rather than optical or electrochemical

Three technologies are available for gas detection in battery systems. Essl and colleagues conclude that metal-oxide sensors are the best candidates for detecting battery failures. The reasons are speed and breadth: MOx is the fastest gas sensor for detecting an incipient thermal runaway and covers a broad band of gases simultaneously.

Form factor matters too. An optical sensor measures around 20 mm in diameter, an electrochemical cell about 9.2 × 12.4 mm. A MOx die measures 1.15 × 1.15 mm. In a battery pack where installation space is counted in millimetres and sensors are meant to sit at several points, that is not a side issue.

Optical methods also bring high power consumption and high price; electrochemical cells age and have to be recalibrated or replaced. Neither sits well with a component that stays in the pack for the life of a vehicle.

The four known weaknesses of MOx — and what became of them

MOx sensors have a poor reputation, for four concrete reasons. Concealing them here would be dishonest; naming them explains at the same time what NANOZ changed.

• Selectivity. The classic objection: a MOx sensor can be misled by pollution events outside the battery pack — false alarms in both directions. This is the point that pattern evaluation across four sensing elements resolves.
• Stability. Ageing and temperature changes shift the baseline. This can be compensated in software if temperature is measured alongside and drift is modelled — and it is why a humidity and temperature sensor belongs on the same board.
• Poisoning. Contamination of the sensitive layer permanently reduces sensitivity. NZGS 2 therefore carries a hydrophobic 0.1 µm filter directly in the package. Silicone compounds remain fatal to the layer nonetheless — that has to be accounted for in the bill of materials.
• Power consumption. A heated sensor draws more current than a temperature or pressure sensor. Around 70 mW at 2 V heater voltage is, however, a figure one can work with in a vehicle.

How four signals become a signature

Four sensing elements and two heaters sit on the same die of 1.15 mm edge length. Because each element has a different layer, a different position relative to the heater and therefore a different local operating temperature, one gas event produces four different signal traces. A deep learning model evaluates those four traces together.

The result is a signature characteristic of every gas concentration and every gas mixture. For the battery application that means: not “some reducing gas is present” but “this pattern corresponds to vaporising electrolyte”. How that works in detail is set out on the technology page; the component's specifications are in the datasheet extract.

One component, many gases — and one algorithm per use case

The division of labour is deliberate. The sensor component — silicon support, transducer, sensitive layer — is the same for an entire group of gases and covers VOCs, carbon monoxide, ozone, ethanol and others simultaneously. The algorithm, by contrast, is tailored to the use case: a single target gas or a specific mixture, identification and concentration, embedded in the customer's microcontroller.

For a battery manufacturer that means the hardware does not need requalifying when the electrolyte formulation changes. What gets retrained is the model.

What legislators now require

Warning time has long since stopped being a matter of product philosophy. With GB 38031-2020 China was the first country to mandate a five-minute warning: after the thermal failure of a cell there must be neither fire nor explosion for at least five minutes, so that occupants can leave the vehicle.

The successor GB 38031-2025 has been mandatory for new vehicles since July 2026 and tightens this considerably: no fire and no explosion for two hours after the initiating event. The alarm signal must be issued no later than five minutes after thermal runaway of a cell is triggered, and no smoke may enter the passenger compartment within that window.

For the sensing side the arithmetic is simple. The five minutes run from triggered runaway. Detect only at first venting and a substantial part of that budget goes on detection itself, leaving correspondingly less for warning, shutdown and intervention. Detect one stage earlier and you gain margin exactly where it is needed.

What the sensor does not do: it does not make a battery system compliant. The tests under GB 38031 address the battery system as a whole, not a component. Earlier detection is one building block of that — not a proof.

Why false alarms cost more than they look

In sensing one distinguishes errors of the first kind — alarm without event — and of the second kind — event without alarm. In the battery application both have unpleasant consequences, but of different character.

A false alarm means, at best, that a vehicle goes to the workshop and nothing is found. At worst the driver learns to ignore the warning — and then the alarm is ineffective when it is justified too. For the manufacturer, warranty cases, recalls and reputational damage follow, the last of which cannot be quantified.

A missed event is the case the standard was written for. It is rarer, but it carries the entire liability.

Both error types arise from the same cause: a sensor delivering only an aggregate value has to decide via a threshold. And any threshold low enough for early detection is also low enough for the interferents of a real environment. Pattern evaluation inverts the question: not “is the value above the threshold” but “does the pattern match vaporising electrolyte”. A passing diesel produces a different pattern regardless of how high its absolute value is.

Why gas arrives before temperature, voltage and pressure

A battery management system measures anyway. The question is what it sees first.

• Cell voltage. An internal short shows as a voltage drop — but only once it is pronounced enough. Early electrolyte decomposition disappears into the noise on the voltage side.
• Temperature. Temperature sensors sit at individual points and measure via thermal conduction. Time passes before local cell heating reaches the nearest probe, and the temperature rise itself starts late — the big jump lies between stage b and stage c.
• Pressure. A pressure sensor in the pack registers venting. That is stage c by definition.
• Gas. Gas travels with the flow inside the pack, not by thermal conduction, and it forms at stages a and b — before there is anything thermal or mechanical to measure.

This is not an argument against the other quantities; a battery management system needs them all. It is an argument that gas measurement is the earliest channel available — and therefore the place where warning time is won.

What a measurement campaign for a cell type requires

For classification to work on a particular cell it has to be trained on it. The effort is manageable but it is not zero, and it cannot be skipped.

What is needed are recordings of the sensor signal under three conditions: reference air for the baseline, the target gases at known concentrations — the electrolyte solvents of the cell type in question, individually and in mixture — and interferents that realistically occur at the installation point. The last is routinely underestimated and decides the false-alarm rate.

Added to that are the boundary conditions: temperature and humidity across the expected range, because both shift the signal, and several specimens of the sensor, because characteristics vary from chip to chip. Only from that data basis does a model emerge that separates cleanly in feature space — and only the cluster separation shows whether the task is solvable at all.

What the sensor does not do in this application

• It does not localise. A sensor reports that electrolyte is vaporising, not which cell is affected. Locating requires several sensors and an evaluation of their arrival-time differences.
• It measures only what reaches it. A pack with sealed module compartments can contain gas for a long time. Flow routing inside the pack co-determines warning time — sensor position is a design topic, not an accessory topic.
• It does not replace redundancy. For safety-relevant functions the usual applies: a single channel is not enough. Use in safety-critical applications requires a separate written agreement under our terms in any case.
• It ages. Baseline shift through ageing is compensable, but it is there. A concept for recalibration or self-diagnosis across the vehicle lifetime belongs in the architecture.

Where the sensor sits in the pack

Two installation positions come into question. Inside the battery pack, at pack level: gas reaches the sensor earliest there, and the environment is thermally and mechanically demanding. Or outside, close to the pack: easier to install and service, but with delay and stronger influence from ambient air.

The decision usually falls in favour of several sensors inside the pack, distributed across the modules — which only works if form factor and power consumption allow it.

How a project runs

From first assessment to series supply, three phases are foreseen. Naming them here saves the question of what getting started actually means.

• Phase 1 — technology assessment and sensor characterisation, about one month. Checking whether the target specification is achievable, and validating the sensitive layer for the use case.
• Phase 2 — AI algorithm development, about two months. Training on the specific gas matrix until proof of use case is established. Development is done by NANOZ or by the customer.
• Phase 3 — integration support, four to six months. Device characterisation and industrial support through to manufacturing.

Customer-side product development runs in parallel, for which six to eighteen months is usual. Between design win and series supply, then, one and a half to two years is realistic — a timeline that fits vehicle projects and that is worth knowing before starting.

What it comes down to

The question in battery monitoring is not whether thermal runaway can be detected — by stage c at the latest, any usable sensor does. The question is whether the report arrives early enough to do something about it, and whether it is reliable enough that someone believes it.

Both depend on the same property: the ability to tell one gas mixture from another. A single-gas sensor does not have it. A MOx sensor with four elements and a classification trained on them does.

Sources: Essl, C. et al., Early Detection of Failing Automotive Batteries Using Gas Sensors, Batteries 7(2), 2021; NANOZ, Detection of Failing Automotive Batteries, September 2023; NANOZ NZGS 2 datasheet, revision 3.3.0.