H₂, CO, CO₂: The Thermal-Runaway Signature

The order matters more than the shortlist. Electrolyte solvent vapour is detectable roughly fifteen minutes before hydrogen, and the gas that dominates the vent depends on the cell: LFP vent gas runs about 41% H₂ against 8% CO, while NMC runs about 36% CO against 19% H₂. A scheme built around a single primary marker is tuned for a cargo a car deck cannot specify.

The signature in order

Four families, in a consistent sequence. The order is what a detection scheme is actually buying — each step is a different chemistry at a different temperature, so each is a separate opportunity with its own lead time.

• Carbonate solvent vapour (DMC, EMC, EC) — first to appear. DMC is a primary component of the electrolyte and its release precedes that of H₂ and CO.
• H₂ — as the solid-electrolyte interphase breaks down, before visible smoke.
• CO and CO₂ — as the organic electrolyte decomposes at higher temperature.
• HF, with POF₃ alongside it — as the fluorinated salt decomposes. This is the toxicity, not the warning.

The gap between the first two steps is the finding that should shape a specification. In one reported test series DMC became detectable at about 390 seconds while H₂ did not appear until about 1,312 seconds — roughly fifteen minutes apart. Purpose-built solvent sensing exists to exploit that: a mixed-phase bismuth oxide DMC sensor has been reported responding to 20 ppm within 6 seconds and raising an alarm about fifteen minutes before runaway onset.

Chemistry inverts the ranking

This is the correction to the version of this post that stood here before, which argued H₂ was the most useful primary without qualifying it by cell. Measured vent-gas compositions do not support a single ranking. LFP releases the least gas, about 0.02 mol/Ah, and it is hydrogen-dominant — roughly 41% H₂, 27% CO₂, 8% CO. NMC releases about 0.07 mol/Ah and is carbon-monoxide-dominant — roughly 36% CO, 24% CO₂, 19% H₂.

Two things follow, and they point in different directions. NMC produces on the order of three and a half times the gas per amp-hour, so the same sensing threshold is reached sooner on an NMC pack for reasons that have nothing to do with how far the event has progressed. And the dominant species flips: on an LFP deck hydrogen is the obvious primary, on an NMC deck it is the third-largest component and carbon monoxide is the obvious one.

And the two channels partially measure each other

Reading both is not the same as reading two independent things. Published cross-interference data for electrochemical cells puts the cross-response between the hydrogen and carbon-monoxide channels at roughly 20–22% — Industrial Scientific describes hydrogen triggering a reading on a CO monitor equal to about 22% of the H₂ concentration present. Because a stressed cell vents both together, neither channel is a clean measurement of its own species during the event it exists to catch.

There is a specifiable answer. Hydrogen-compensated CO sensors exist precisely for this: they measure H₂ and subtract it from the combined reading, and some designs use two electrodes and a catalyst with lower hydrogen sensitivity. If a scheme leads on CO — which the NMC composition argues for — that compensation belongs in the specification rather than in a footnote.

What HF actually costs

HF closes the sequence, and it is worth quantifying rather than calling it dangerous. Measured emissions run between 20 and 200 mg of hydrogen fluoride per watt-hour of nominal cell energy, with 15–22 mg/Wh of phosphoryl fluoride recorded alongside it in some tests. Extrapolated to a vehicle-scale pack, that is on the order of 2 to 20 kg of HF from a single 100 kWh battery.

Set that against the exposure limits. The IDLH level for HF is 30 ppm, and the ten-minute lethal value is around 170 ppm. In an enclosed vehicle deck those are not large numbers against kilogram-scale release. HF is therefore not a detection opportunity at all — by the time it is present the useful decisions have already been made or missed. It is the reason the earlier markers are worth instrumenting.

Where gas sensing still fails on a deck

Two constraints bound everything above, and neither is a sensor problem. The first is that the source is sealed: a traction pack is built gas-tight and releases through a calibrated burst path, so nothing reaches any sensor until the pack decides to vent — the signature's clock starts at the aperture, not at the cell. The second is the space: EMSA's FIRESAFE II work identified ventilation on open ro-ro and weather decks as capable of both delaying detection and delocalising it, so an alarm can arrive late and point away from the source.

That is why the gas layer is worth building and why it cannot stand alone. Gas sensing tells you the deck has a problem. It does not tell you which vehicle. Localisation has to come from somewhere else, and the ratio between channels — rather than any single threshold — is what carries the diagnostic information once the plume has mixed.

Sources

• 1. Vent-gas composition by chemistry: LFP releasing approximately 0.02 mol/Ah at roughly 41% H₂, 27% CO₂ and 8% CO; NMC approximately 0.07 mol/Ah at roughly 36% CO, 24% CO₂ and 19% H₂. Principal components across chemistries reported as CO, H₂, CO₂ and preserved electrolyte solvents, with minor CH₄, C₂H₄, C₄H₁₀ and H₂O. A separate NMC series reports H₂ at 36.0 vol%, CO₂ at 26.7 vol% and CO at 23.8 vol%. Drawn from the comparative vent-gas literature including Batteries (MDPI) 2019, 5(3):61 and Batteries (MDPI) 2025, 11(9):323 (NVPF sodium-ion versus NMC and LFP).
• 2. Sequence and lead time: DMC as a primary electrolyte component whose release precedes H₂ and CO; a reported series in which DMC became detectable at ~390 s and H₂ at ~1,312 s.
• 3. Solvent sensing: a mixed-phase bismuth oxide DMC sensor reported responding to 20 ppm DMC within 6 seconds and triggering an alarm approximately 15 minutes before the onset of thermal runaway — Sensors and Actuators B: Chemical (ScienceDirect S0925400526003291). Related work on H₂/CO sensor arrays with signal processing for early runaway detection at ScienceDirect S0925400525021148.
• 4. Cross-interference between the H₂ and CO channels: Industrial Scientific's electrochemical sensor cross-interference material describes hydrogen producing a reading on a CO monitor equal to about 22% of the H₂ concentration, and notes hydrogen-compensated CO sensors that measure and subtract H₂, plus two-electrode designs using a catalyst with lower hydrogen sensitivity.
• 5. HF and POF₃ yield and toxicity: 20–200 mg/Wh of hydrogen fluoride and 15–22 mg/Wh of phosphoryl fluoride measured in Li-ion fire tests, extrapolating to roughly 2–20 kg of HF for a 100 kWh pack; HF IDLH given as 30 ppm (0.025 g/m³) and the ten-minute AEGL-3 value as 170 ppm (0.0139 g/m³). Larsson et al., "Toxic fluoride gas emissions from lithium-ion battery fires", Scientific Reports (2017) 7:10018.
• 6. Deck-scale constraints: EMSA FIRESAFE II identified ventilation in open ro-ro spaces and weather decks as able to both delay and delocalise detection — covered in this corpus at FIRESAFE II: Open-Deck Detection Lessons. The sealed-pack egress constraint is set out at Does the Off-Gas Ever Leave the Battery Pack?.