First, understand precisely what ‘decisive’ means

In battery-fire detection, carbon dioxide (CO₂) is a highly important signal. It does not, however, mean that observing CO₂ alone will ensure that no fault is missed. In this article, ‘decisive’ means a strong indication that electrolyte decomposition and side reactions have begun inside the cell before thermal runaway becomes visible as a fully developed flame. In actual installations, signal magnitude and arrival time vary with cathode chemistry, cell format, state of charge (SOC), ageing, fault-initiation mode, enclosure volume, ventilation rate, and sensor location. Background CO₂ is also produced by people and combustion equipment. A sound system therefore does not declare a fire solely from the absolute CO₂ concentration; it interprets the increase above baseline and the rate of rise together with information from other sensors.

This distinction matters in practice. A smoke detector requires combustion particles to travel to the detector, while a surface-temperature sensor may see reactions that began inside the cell only later. Gas, by contrast, may be released when the cell vent opens before a visible flame appears. CO₂ detection is a layer that can capture this short lead interval and provide time to begin actions such as stopping charging, isolation, ventilation control, and on-site verification. However, a CO₂ sensor may also remain silent where an internal short circuit does not open the vent or where the gas cannot reach the sensor.

What happens inside a battery before thermal runaway

A lithium-ion cell consists of a cathode, an anode, a separator, electrolyte, and a sealed case. When internal temperature rises because of overcharge, external heating, an internal short circuit, impact, or a manufacturing defect, protective interphases and electrolyte decompose and multiple exothermic reactions proceed in sequence. This process generates CO₂, carbon monoxide (CO), hydrogen (H₂), hydrocarbons such as methane (CH₄) and ethylene, and volatile organic compounds (VOCs). When internal pressure exceeds the design limit, the safety vent opens and a mixture of gases and vapours is expelled. Even without an ignition source, flammable components can accumulate in an enclosed space and create a risk of delayed ignition and deflagration.

In tests by the U.S. Federal Aviation Administration (FAA) of lithium cobalt oxide cells at several SOCs and heating rates, the principal components of collected vent gas included CO₂, H₂, and CO. In both the pouch and cylindrical cells tested, total vent-gas volume and combustion energy increased with SOC. However, the proportion of CO₂ in the gas tended to decrease as SOC rose, while the proportion of CO increased. The key lesson is that ‘more CO₂’ and ‘greater hazard’ are not always equivalent statements. CO₂ is a useful indicator of abnormal reactions, but flammability, toxicity, and explosion hazards must also be assessed with H₂, CO, hydrocarbons, and oxygen concentration.

Three reasons CO₂ is advantageous for early detection

First, CO₂ is repeatedly observed when many types of lithium-ion cells undergo thermal or electrical abuse. In FAA tests, CO₂ represented a substantial share of vent gas from the test cells, and experimental research comparing NCM and LFP cells also assessed gas emissions including CO₂ as indicators of thermal-runaway risk. This means CO₂ is not an incidental by-product appearing only in one model from a particular manufacturer; it is linked to decomposition reactions of the electrolyte and electrode interfaces. It does not mean, however, that every chemistry and failure mode produces the same concentration curve.

Second, CO₂ can be measured relatively selectively by non-dispersive infrared (NDIR) methods. CO₂ molecules absorb infrared radiation near 4.26 micrometres. An NDIR sensor measures how much infrared radiation emitted by a source is reduced at a specified wavelength after passing through a gas cell. Because it uses a target wavelength rather than relying solely on reactive consumables as electrochemical sensors do, it can be designed for long-term monitoring. Dust, condensation, optical-window contamination, changes in temperature and pressure, and long-term drift nevertheless still require correction and maintenance.

Third, CO₂ may capture venting before visible smoke or flame. In an overcharge experiment with a 50 Ah LFP cell, researchers monitored CO₂ and CH₄ by NDIR and detected the target gases about 25 seconds before the sharp voltage drop and thermal runaway under closed-vent conditions. Under conditions in which the vent was opened in advance, they detected gases about 580 seconds before the battery surface reached the usual 60°C temperature alarm. These figures are results of that test configuration, not response times guaranteed for every ESS. They nevertheless illustrate why monitoring the gas pathway as well as surface temperature is valuable.

Why CO₂ cannot be the sole answer

CO₂ is already present in a battery room. Outdoor air contains CO₂ at roughly several hundred ppm, and concentrations may be higher indoors where people are present or around combustion equipment. Sensor readings can change simply because the number of people entering changes or HVAC operation is altered. A fixed threshold applied to one absolute value can increase false alarms or, conversely, miss a small battery signal against a high background concentration. It is therefore preferable to establish baselines by normal operating time period and consider both short-term change and rate of rise. A differential approach comparing values near the outdoor-air intake and battery racks can also help distinguish background changes from local releases.

Ventilation changes the signal substantially. Strong airflow dilutes gas, lowering the concentration at a sensor, and can carry it to the opposite side of the sensor. In a sealed cabinet, even a small release accumulates rapidly, whereas the same release disperses at a lower concentration in a large open space. Placing only one sensor on the ceiling may not be sufficient. Although CO₂ itself is heavier than air, battery vent gas is hot and mixed with light components such as H₂, so it initially follows buoyancy and airflow. Rather than simply choosing either the floor or ceiling, the design must account for the actual vent direction, rack ducts, cooling fans, exhaust outlets, and stagnant zones.

Cell chemistry and condition also differ. NCM, NCA, LCO, and LFP do not have the same reaction temperatures or release compositions, and results vary even within the same chemistry with SOC, cell size, case construction, and ageing condition. The FAA report likewise explains that thermal-runaway outcomes depend on chemistry, size, SOC, and manufacturer design and configuration. It is therefore hazardous to copy a ppm threshold obtained from another product directly. On-site alarm values should be validated against cell-, module-, and rack-level test data for the installed product and manufacturer information, and, where possible, staged test results such as UL 9540A.

A more reliable alarm through multiple sensors

The purpose of multiple sensors is not simply to fit many sensors, but to overlap different failure windows. The BMS continuously monitors cell voltage, pack current, SOC, insulation, and some temperatures. Temperature sensors can observe local heat generation and cooling abnormalities; smoke sensors, aerosols and combustion products; and pressure sensors, rapid changes in sealed enclosures. CO₂ and VOCs provide chemical traces of venting; CO contributes to assessment of incomplete oxidation and toxic hazard; and H₂ or combustible-gas sensors contribute to assessing explosion potential. Oxygen sensors can confirm inerting or oxygen-deficient conditions. No single one responds first in every situation.

Experiments with commercial diagnostic technologies at Sandia National Laboratories demonstrate this point well. Researchers compared electrochemical impedance (EIS), gas sensors, voltage, and temperature together. In some overheating tests, EIS indicated an abnormality first; in overcharge pack tests, they tested an intervention in which a combined gas sensor detected off-gas and then cut current to prevent progression to thermal runaway. By contrast, gas sensors did not respond in tests where cells did not vent. Thus, electrical abnormalities require complementary BMS and impedance layers, venting a gas layer, and heat generation and combustion temperature and smoke layers.

Full-scale ESS tests by FSRI and UL likewise measured temperature, pressure, O₂, CO, H₂, HCN, VOCs, and combustible-gas indicators together. The researchers noted that fire services need early access to internal ESS measurement data, particularly gas information. They also cautioned that portable gas meters and thermal-imaging cameras have limitations, and that ventilation itself can move an accumulated flammable mixture into its ignitable range, causing deflagration or rapid fire spread. Accordingly, rather than simple automation such as ‘turn the fans to maximum when gas is detected,’ the design must combine concentration, flammability range, ignition-source isolation, and equipment-specific emergency procedures.

Staged decision logic for field application

Operational alarms can be staged as attention, abnormality, and emergency. At the first stage, data are checked when CO₂ rises by a defined amount above baseline or its rate of rise is abnormal. Rather than immediately initiating automatic suppression or full evacuation, adjacent sensors, BMS events, and HVAC status are cross-checked. At the second stage, charge and discharge are safely stopped and the affected rack isolated when the CO₂ increase persists together with a rise in one of VOCs, CO, or H₂, or with cell-voltage deviation, temperature rise, or insulation abnormality. At the third stage, the designed emergency shutdown, evacuation, notification of fire services, and access control are executed when rapid changes in temperature or pressure, smoke, rising combustible gas, or signs of spread to multiple racks appear.

Insisting only on AND logic can miss a real incident because one sensor fails, while relying only on OR logic produces frequent false alarms. A more practical approach weights the reliability and risk level of each sensor. For example, a very rapid CO₂ rise and a voltage abnormality in the same rack receive a high score, whereas slowly changing indoor CO₂ alone receives a low score. Sensor self-diagnostic failures and communication loss should also be included in the risk score. After an alarm, retain raw trends, ventilation status, BMS events, and time-synchronised records for root-cause analysis and threshold improvement.

Placement and commissioning matter more than sensor specifications

Even a good sensor delays early warning if installed in the wrong place. Designers should first confirm the cell-vent direction and internal flow paths of the module. Next, they should verify where normal cooling airflow and emergency exhaust airflow carry gas, using smoke visualisation or tracer-gas testing. Sensors can be layered at anticipated collection points inside racks or cabinets, in common exhaust ducts, and at representative indoor points. Rather than observing concentration at one point only, arrival-time differences and spatial gradients among sensors can narrow down the affected rack more quickly.

During commissioning, record background values under actual operating conditions, including normal load, maximum ventilation, ventilation shutdown, and changes in the number of people entering. Sensor response time depends not only on the sensing-element specification but also on sampling-tube length, pump flow rate, filters, and enclosure leakage. Calibration by holding test gas directly in front of the sensor cannot validate system response. Where possible, inject a safe tracer gas at the anticipated release point and measure the entire time from detection, transmission, and alarm to operator confirmation. If test results do not meet the target lead time, adjust location, sampling flow rate, and logic.

Maintenance and false-alarm management

Although CO₂ NDIR sensors are comparatively stable, they are not maintenance-free devices. At intervals specified by the manufacturer, verify zero and span and inspect optical contamination and condensation, pumps and filters, and tube leakage. A sensor that appears fixed at a normal concentration is not necessarily safe. If its value is excessively flat over a long period or inconsistent with adjacent sensors, the cause may be blockage, communication error, or calibration failure. Alarm-bypass status and calibration due dates should be clearly visible on the operations dashboard.

False alarms are not data to discard; they are data from which the design can learn. Recording which activities affected which sensors—such as workers’ breathing, forklift exhaust, cleaning agents, welding, and HVAC switching—can refine alarm logic. However, repeated false alarms must not be used to raise thresholds without evidence. First separate the background cause and improve sensor location and cross-validation conditions, then compare before and after changes through simulated testing. When an alarm sounds, the operator display should show not only the current value but also the recent rate of rise, baseline, related BMS alarms, sensor status, and recommended actions.

Design CO₂ as a layer, not as a single vote

The real value of CO₂ is not in confirming its presence after a fire has grown. It lies in detecting the moment when a cell begins abnormal reactions and vents, thereby advancing the next action. In multiple experiments, CO₂ was observed as a principal vent gas, and NDIR sensors provided a selective, continuous means of measurement. CO₂ is therefore a signal too valuable to omit from early-warning design for battery rooms and ESS cabinets.

However, a decision must not be made by one CO₂ number. Chemistry-specific release characteristics, SOC, ventilation, background concentration, and sensor location must be reflected through field testing and combined on a time axis with voltage, current, temperature, VOC, CO, H₂, smoke, and pressure information. The best system is not one in which a sensor ‘guesses’ the incident; it is one that progressively turns uncertain early signals into confidence and provides people time to intervene safely. CO₂ may be an important first voice in that decision network, but it must always be heard with the others.

References

Cover photo

Original photo from Pexels