Failure Modes and Thermal Runaway: Heat, Gas, Fire and Explosion

Most battery failures start small: one cell, one connection, one leak. Whether that ends as a smoking cabinet, a container that burns for hours or a blast that throws firefighters across a yard depends on what happens in the next few minutes inside the enclosure. That chain is what codes, designs and fire tactics are trying to break.

What sets a cell off

Inside thermal runaway

Sandia's handbook describes the core problem: a cell operated outside its limits can make heat faster than it can shed it, and once a metal-oxide cathode gets hot it releases oxygen while the organic electrolyte supplies fuel. The steps below are simplified; real cells overlap them.

Propagation is the step that turns one bad cell into a plant fire. Module barriers try to keep heat inside one cell or module. Spacing and walls try to keep a fire inside one rack or container. Codes now ask manufacturers to prove, by test, how far a failure spreads; lesson 8 covers UL 9540A and large-scale fire testing.

What comes out

Failing cells vent a mix of gases and particles. The 2024 Journal of Energy Storage review of 60 studies found the average gas mix similar across LFP and NMC, with more hydrogen from LFP and more carbon monoxide from NMC. Larsson and colleagues, in a 2017 Scientific Reports study burning seven types of commercial lithium-ion cells, measured hydrogen fluoride at 20 to 200 milligrams per watt-hour of battery capacity, plus 15 to 22 mg/Wh of phosphoryl fluoride in some tests. The Arizona hazmat team in 2019 measured dangerous levels of hydrogen cyanide and carbon monoxide inside the container's hot zone, per UL FSRI.

Fire or explosion

When vented gas ignites as it comes out, the result is a fire: hot, smoky, long, but without a pressure blast. When gas pours into a closed enclosure and does not ignite right away, it mixes with air. If the mix reaches its flammable range and then finds a spark, hot surface or flame, it deflagrates. That is the pattern in the worst incidents in EPRI's database.

• Surprise, Arizona, April 19, 2019. Clean agent discharged; gas built for about three hours; the door was opened and about three minutes later the container deflagrated. Four firefighters seriously injured.
• Beijing, April 16, 2021. A 25 MWh LFP system exploded; two firefighters died and a third was injured, with one employee unaccounted for.
• Aghione, Corsica, April 6, 2023. Shortly after doors were opened, an explosion injured two firefighters; three more were hospitalized for smoke.
• Isa, Kagoshima, Japan, March 27, 2024. Firefighters opened the door and the explosion came when they tried to use the smoke exhaust system; four were injured.
• Gunwi, South Korea, January 17, 2022. An explosion occurred ten minutes after firefighters left the structure; no injuries.

Of the 165 rows in EPRI's stationary table as of September 2026, 19 are flagged as explosions. Most incidents are fires without a blast, and EPRI's October 2026 report found 93 percent of recorded utility-scale incidents caused no injuries or deaths, with no reported U.S. BESS injuries or deaths since the 2019 Arizona event.

How designs control explosions

Prevention keeps gas below its flammable range with detection and exhaust ventilation; NFPA 69 is the standard. Venting lets the blast happen safely: panels that open under pressure and point the force upward, away from people and other units; NFPA 68 is the standard. Many systems use both. Sandia notes that active suppression can make gas buildup worse, because gas that is not burning accumulates, which is why gas detectors for carbon monoxide and hydrogen have been added. The 2026 NFPA 855 adds a unit-level test with intentional ignition of vent gases when cell or module tests release flammable gas.

Off-gas volume that reaches the LFL:
  LFP: 60 m3 x 0.062 = 3.72 m3
  NMC: 60 m3 x 0.079 = 4.74 m3

A ventilation design limited to 25% of the LFL
(the 2018 IFC ventilation target):
  LFP: 3.72 x 0.25 = 0.93 m3
  NMC: 4.74 x 0.25 = 1.19 m3

Ratio: 3.72 / 4.74 = 0.78
From the LFL alone, LFP gas makes this box
flammable with about 22% less gas than NMC gas.

Why systems fail: the root cause data

EPRI, TWAICE and Pacific Northwest National Laboratory classified incidents in EPRI's database by root cause (design; manufacturing; integration, assembly and construction; operation) and by failed element (cell or module, controls, balance of system). Their May 2024 white paper found:

• Only 26 of 81 incidents had enough public information to classify.
• Integration, assembly and construction was the most common root cause, about 10 of the 26 by Utility Dive's count. Manufacturing was the least common.
• Cells and modules were the failed element in 3 of the 26; controls and balance-of-system parts made up most of the rest.
• 72 percent of failures with known system age happened during construction, commissioning or the first two years of operation.

The incident record fits that picture: the Moorabool, Australia fire happened during commissioning; the Tilbury, England fire in February 2025 happened while the site was still under construction; the two Warwick, New York failures in June 2023 hit new systems during a large storm. EPRI's October 2026 report added that failure rates fall with system age and with newer vintages, and recommended safety reviews and retrofits of early-vintage systems.

The data has a real limit: two-thirds of incidents could not be classified because the public record was too thin. Neighbors of a failed site often never learn the cause.