Thermal Runaway Propagation: How One Cell Takes a Pack

One lithium-ion cell in runaway gives off very little energy. Tens of kJ, or about the heat from a gram of petrol. So the disaster never comes from that one cell. Thermal runaway propagation brings it, since the failed cell drives its neighbours past their own trigger point. The fault then walks through the module, into the pack, along the rack, and at grid scale through a whole box. This post covers the physics of that walk, and the design that halts it.

TL;DR

How thermal runaway propagation works

• The cascade follows a heat budget. A failed cell gives off some 0.3–1.2 kJ per Wh of stored energy, and just 20–30% of that leaves through the can.
• In a module where cells touch, about 12% of one cell’s heat can trigger the cell beside it. Good design must block nearly 88% of the path.
• Four paths carry the heat: conduction through the can, radiation, hot vent gas, and molten spray.
• Cell-to-cell steps cluster near 140–250 s for big cells at full charge. Tight NMC stacks can drop to single digits.
• Charge state drives the violence. One 50 Ah LFP cell makes 0.90 mol of gas at 50% SOC, yet 2.33 mol at 100%.

What stops thermal runaway propagation

• Passive design wins. Aerogel or mica sheets of 1–3 mm, air gaps, metal spreaders, and vents aimed away from neighbours cut the heat that lands next door.
• Water and clean agents cannot halt runaway inside a sealed module. They cool the second fire and guard the next rack instead.
• Gas control saves lives. At McMicken the blast came three hours after the first cell failed.
• China now writes the goal into law. GB 38031-2025 asks for no fire and no blast for 120 minutes after one cell fails, from 1 July 2026.

What does thermal runaway propagation mean?

Thermal runaway propagation means the spread of runaway from cell to cell inside a pack. One cell fails, heats the cell next to it past its trigger point, and the chain repeats.

No step after the first one needs a new flaw. That first cell may fail from an internal short, a nail, an overcharge, or a leak of coolant that then arcs. After that, layout and heat flow decide the rest. So a pack with one bad cell either loses one cell or loses the lot.

Why does one bad cell matter so much?

Because a pack stores its fuel in a tight lattice, and each neighbour sits millimetres from the heat.

Look at the sums. One 3.5 Ah 18650 cell gives off about 60 kJ at full charge. A 5 MWh box holds 18 GJ. Nothing in that box needs an outside spark once the first cell vents, since the pack brings both fuel and heat.

Which three temperatures decide a cell’s fate?

Three points frame every runaway: T1 for the onset of self-heating, T2 for the trigger, and T3 for the peak.

According to the scheme that Feng Xuning and Minggao Ouyang’s group at Tsinghua drew from calorimetry, T1 marks the first steady exotherm, near 80–120 °C. The SEI layer breaks down there, and the anode starts to react with the electrolyte. T2 marks the point where self-heating tops about 1 °C per second and nothing stops the climb. In hardware terms that means a collapsed separator and a huge internal short, near 200–260 °C for many NMC cells. T3 marks the peak: about 500 °C for LFP, and 800–1000 °C or more for high-nickel NMC.

The gap from T2 to T3 tracks how bad the event gets, and how much heat the cells next door must absorb.

Which cell’s trigger point counts?

The one that receives the heat. A failed cell either drags its neighbour up to that neighbour’s own T2, or fails to.

Nobody can soften the runaway of a cell already gone. Yet you can starve the next cell of the heat it needs, and that lever stays live right up to the moment a barrier melts.

What sets the heat budget for thermal runaway propagation?

One balance sets it. The cascade goes on while the heat delivered beats the heat a neighbour needs to reach its own trigger point.

Treat the module as a network of thermal resistances. For a cell of mass $m$ and specific heat $c_p$, the heat needed reads:

$Q_{\text{needed}} = m\,c_p\,(T_2 - T_{\text{amb}})$

The heat delivered before the failed cell burns out reads:

$Q_{\text{delivered}} = \int_0^{\tau}\Big[\underbrace{\tfrac{T_{\text{TR}}-T_{\text{cell}}}{R_{\text{cond}}}}_{\text{casing/busbar conduction}} + \underbrace{\varepsilon\sigma A\,(T_{\text{TR}}^4 - T_{\text{cell}}^4)}_{\text{radiation}} + \underbrace{\dot m_{g}c_{p,g}(T_g-T_{\text{cell}})}_{\text{hot-gas + jet-flame convection}} + \dot q_{\text{ejecta}}\Big]dt$

The cascade runs while $Q_{\text{delivered}} > Q_{\text{needed}}$. Every product on the market attacks one term there. Sheets raise $R_{\text{cond}}$, while gaps cut the view factor in the radiant term. Heat sinks pull energy out of the target cell. Aimed vents strike the gas and spray terms from the neighbour’s account.

Which heat path dominates?

Contact through the can leads in tight prismatic and pouch modules. Radiation and flame jets take over once cells sit apart.

In Feng’s 25 Ah NMC module tests, the can carried some ten times the heat of the pole connectors. A study of round cells found the reverse for cells with no metal link to the failed one, where radiation led. One smoke study put 80.5% of the transport down to contact in a sealed rig, yet the hot smoke still cut the step time by 44.9%.

Why does only 12% of the heat matter?

Because the next cell needs far less heat to reach T2 than a full runaway gives off. Feng’s group put the share near 12% for 25 Ah NMC cells in contact.

That number explains the pain. To stop the chain you must block some 88% of the path, and do so while the sheet itself faces a surface near 800 °C. Read the figure as layout-bound, though, not as a law of nature. One review of NCM modules cites about 10%, and both numbers come from flat cells pressed together.

Where does the heat actually go?

Most of it leaves with the spray. The can carries just a fifth to a third of the total.

Fractional runaway calorimetry, built by NASA with NREL, gives the split. For 18650 cells at full charge, the median total lands near 60 kJ, with 20–30% out through the can and 70–80% out with gas and hot solids. One LG 18650-MJ1 cell gave 59.6 kJ, split 18% through the can, 79.8% out of the positive end, and 2.2% out of the negative end.

How you trigger the cell shifts that split, which matters for test fairness. A nail pushes about 34% of the heat through the body, against some 21% for a heater. Two labs can thus test one cell design, report two different outcomes, and both stay right.

Does chemistry change the heat budget?

Yes, and the ranking flips with the yardstick you pick.

Per Wh of stored energy, ARC data give some 0.9–1.2 kJ/Wh for LFP and LTO, against just 0.2–0.3 kJ/Wh for NMC. LFP thus gives off more heat per unit of stored charge. Yet NMC packs far more charge into the same space, and it peaks hundreds of degrees hotter. Sandia’s work on energy density shows total runaway heat rising near linearly with charge state, while the peak rate of self-heating climbs steeply at the top. That rate, rather than the joule count, drives the split between the two chemistries. Our post on LFP vs NMC walks through the breakdown steps behind those curves.

How fast does thermal runaway propagation travel?

Between about 140 s and 250 s per cell for large cells at full charge. Tight NMC stacks can fall to single-digit seconds.

Feng’s six-cell 25 Ah NMC module gave steps of 137–186 s. The first step ran longer, 210–481 s, because the nail acted as a heat sink and stole energy from cell one. Discount that first step when you read any test report.

Theory now backs the stopwatch. Feng’s group solved for the speed of the front and found a half-power link between thermal conductivity and reaction rate. The same work predicts a frozen front, that is, a chain that stalls once heat loss plus transfer delay beat heat gain.

Does a lower charge state stop the spread?

Less charge slows thermal runaway propagation every time, yet no fixed cut-off exists.

Side-by-side module tests from FSRI’s task group work showed a full module running faster and harder than a twin at 30%. Gas output tells the same tale with firmer numbers. One 50 Ah LFP cell gave 0.68–0.90 mol of gas below 50% SOC, then 1.18 mol at 75% and 2.33 mol at 100%. Hydrogen rose from 0.01 mol when flat to 0.93 mol when full. So the lower flammable limit of the vent gas fell from 55.4% to 6.9%, which turns a bad smell into a bomb.

Where the cut-off sits depends on heating rate, cell format and spacing. Sandia’s pouch-stack tests found the safe boundary sliding to lower charge as the heating rate rose. In dense 18650 packs, half charge slowed thermal runaway propagation without breaking it. Some stacked round-cell layouts do hold at 50%, and module models point to a stop below about 30%. Treat charge limits as depth of defence, never as proof.

Does LFP resist the spread better than NMC?

Yes, by a wide margin in car cells. Roughly five times slower across a whole module.

A test of real car cells put NMC-811 from a Mini Cooper SE against LFP from a Tesla Model 3. Reaction speed ran nine times faster in the NMC cells, and the whole-module step ran five times faster. The LFP module often failed to spread at all inside the five-minute window that the old GB 38031-2020 rule allowed. Ejected mass follows the same order, at some 20–25% of cell mass for LFP against 40–50% for NMC.

Format counts as much as chemistry. Big round cells throw more heat per Ah. NASA also found that thin-walled cells above 660 Wh/L tend to split at the side wall, which defeats aimed venting and hands the fire straight to the next cell.

How do engineers stop thermal runaway propagation?

By attacking one side of the balance. Cut the heat delivered, raise the heat drained, or lift the trigger point of the next cell.

Real packs blend several levers, and each one costs energy density.

Do air gaps alone work?

For mild cells, sometimes. For dense NMC, rarely.

Gaps attack contact and radiation at once, and models of round-cell modules confirm that wider gaps, smaller solder joints and thinner tabs all slow thermal runaway propagation. Radiant coupling stays strong once a neighbour glows, though, since that term climbs with the fourth power of temperature.

Which barriers earn their place?

Thin aerogel and mica sheets, mostly from 1 mm to 3 mm.

One odd lesson runs through that table. Pure insulation can backfire, since it traps heat in the failed cell and makes that cell’s own runaway worse. Above a certain specific energy, the field now calls for a sheet plus a heat path to halt thermal runaway propagation.

Can a heat sink beat a barrier?

Sometimes it can. A metal spacer touching every cell drains heat faster than the neighbours can bank it.

NASA showed modules that resist the chain with no active parts above 190 Wh/kg, using aluminium spacers and heat-pipe spines.

One caution belongs beside that win. The cooling loop can start the fire itself, as the Victorian Big Battery showed when a coolant leak drove arcing in the power electronics.

Where should the vent gas go?

Away from every neighbour cell, then out of the box through a vent sized for the job.

Aimed venting strips the 70–80% of runaway heat that rides with the spray out of the neighbour’s budget. NASA’s test work concluded that no spread, with no flames or sparks, stays within reach at a small mass cost. One condition applies: the design must shield nearby cells from the vent path and stop side-wall tears. At box scale the same idea becomes blast venting under NFPA 68 or EN 14994, sized so trapped gas never reaches a blast mix. NFPA 69 offers the other road, holding the air below 25% of the lower flammable limit by inert gas or by fans.

Does suppression stop runaway inside a module?

No. Water and clean agents cannot reach the reacting jelly roll, and the cathode makes its own oxygen.

At McMicken the clean-agent system dumped as designed and still failed to stop thermal runaway propagation. What suppression does well, it does outside the module: cooling the second fire, wetting the next rack, and buying time for crews. FM Global’s tests show good sprinklers slowing or stopping spread to nearby racks. Read the scope with care. Rack to rack, yes. Inside a sealed module, no.

How far apart must racks sit?

NFPA 855 starts at 3 ft, or 0.9 m, between units and to walls, with a 50 kWh unit cap and a 600 kWh group cap.

Those numbers hit modern hardware head on. A 20-foot box now holds around 5 MWh, some eight times the group cap. So almost every utility job takes the waiver route, using large-scale UL 9540A data to show no thermal runaway propagation for the local fire chief. Insurers often ask for 8 ft or more all the same.

Do cell-level guards help?

They cut the electric share of the fault, and nothing else.

Current-interrupt discs, PTC rings, safety vents and fuse tabs all act before or during the short. Busbar fuse links stop a parallel group from dumping its charge through the failed cell. Yet none of them touch a heat path once the cell burns, which explains why packs still need sheets and gaps.

What do the standards ask for?

Rules split along a clean line. China writes the outcome into law, while the USA and the EU write the process.

What does UL 9540A actually certify?

Nothing, strictly. UL 9540A gives a stepped test method and a data pack, never a pass mark.

The method climbs four rungs: cell, module, unit, site. Cell tests force runaway in a vessel and log the vent gas, its flammable limit, peak pressure and burn speed. Module tests force one or more cells into runaway and watch for thermal runaway propagation. The 2025 edition also adds a hard site case that assumes a blast has already happened.

Four critiques belong in any spec. First, results swing with the trigger method, as the nail-versus-heater split above shows. Second, a module test that wants thermal runaway propagation shown can reward a weaker module. Third, FM Global’s DS 5-33 states plainly that results do not stretch across chemistries, formats or racks. Fourth, a pass at day one says nothing about an aged cell. FM’s own UPS tests found a system with a UL 9540A pass still spreading to nearby modules, with peak heat release near 700 kW and module skins above 260 °C.

Why does GB 38031-2025 set the toughest bar?

Because it fixes the outcome. No fire and no blast for at least 120 minutes after one cell fails.

Published in March 2025, the rule bites on 1 July 2026 for new type approvals and 1 July 2027 for models already on sale. The 120-minute clock replaces a five-minute warning window, a 24-fold jump. Smoke must also stay harmless to the people on board. The revision adds 7 cell tests and 17 pack tests, among them a new bottom-impact test and a safety test after 300 fast charge cycles. Chinese makers began to claim certified cells and packs within weeks, so the target has already reached the production line.

How do the USA and the EU compare?

Both write the process rather than the outcome. Assess, document, warn.

The US rule FMVSS 305a tracks UN GTR 20 with a lighter hand, since NHTSA judged single-cell detection too design-restrictive. UN GTR 20 Phase 1 asks for a five-minute escape warning after one cell fails, while the Phase 2 test method for thermal runaway propagation remains in the works. Europe works through EN IEC 62619 and IEC 62933-5-2, with Regulation 2023/1542 wrapping safety into a lifecycle frame. None of those texts yet says what GB 38031-2025 says.

Which incidents prove the point?

Four events shaped current practice. Each one turned on thermal runaway propagation and on design detail, not on exotic cell flaws.

APS McMicken, 19 April 2019

A 2 MW/2 MWh site in Arizona gave the field its textbook case. DNV GL traced the start to an internal fault in one cell pair, likely from odd deposits of lithium metal, which then ran through the rack. The named factors read like a checklist for this post: no sheets between cells, a clean-agent system that could not stop runaway, trapped off-gas, and no plan with the fire service.

About three hours later, crews opened the door. Fresh air met the trapped gas, and the blast hurt the crew. Counts differ across reports: coverage of the DNV GL work cites eight firefighters and one police officer, while UL’s Fire Safety Research Institute title cites four firefighters. LG Chem’s own experts also rejected the cell-fault finding and proposed outside arcing.

Victorian Big Battery, 30 July 2021

Two Tesla Megapacks burned during commissioning of a 300 MW/450 MWh site in Australia. The team traced the root cause to a coolant leak that arced in the power electronics and drove cells into runaway. Fire then reached the next unit as wind-driven flames lit its plastic vent covers, despite correct spacing. Winds near 35 mph that day beat the 12 mph envelope behind the layout.

Just 2 of 212 Megapacks took damage. Unit spacing thus held as a site-scale firebreak even while it failed between two neighbours. Crews had also switched off safety features by keylock during commissioning, a reminder that guards protect nobody while off.

Moss Landing, 16 January 2025

The 300 MW Phase 1 building at Vistra’s California site burned, with EPA putting the loss near 55% of some 100,000 cells. Nobody died, about 1,200 residents left their homes, and later work put some 25 metric tons of heavy metals across nearby wetlands.

Two lessons stand out. Inside the building, thermal runaway propagation ran nearly to the end through a dense indoor rack layout that few would build today. Between buildings, the firebreak held, so the site kept its other storage halls. Flare-ups went on for weeks, and the full root-cause report stayed open at the time of writing, so treat the damage share and the suppression theory as early findings.

South Korea, 2017 to 2019

Up to 28 storage fires hit the Korean market in about two years, taking some 35% of units offline and costing more than $32 million. The state review named four causes: weak guards against shock, poor site conditions, sloppy installs, and control-system gaps. A later phase found that fires clustered above 95% SOC, during or just after a full charge.

Cell makers rejected that finding, as at McMicken. Even so, the run of fires remains the clearest large sample we have that running and wiring, not cell making, drive real losses.

Does the record show progress?

Yes, and sharply so. Failures per GW deployed fell about 97% between 2018 and 2023.

EPRI’s public database, built with PNNL and TWAICE, puts the rate near 9.2 failures per GW in 2018 against about 0.2 in 2023. Fleet size over that span grew from about 1 GW to 65 GW. Of the sorted cases, just 11% trace to a cell or module, while some 65% trace to running and wiring. Both numbers carry the thesis here: the cascade yields to design, and the design already works.

How should you design against thermal runaway propagation?

Start at the module, never at the fence line. Eight habits then cover most of the ground.

For grid-scale designers

1. Aim to stop thermal runaway propagation at module level first. Treat unit and rack spacing as the second firebreak, not the first. Pass UL 9540A at module or unit level, and the site-scale case never arises.
2. Prefer LFP for stationary storage unless volume forces NMC. Five times slower thermal runaway propagation, and a peak hundreds of degrees lower widen every margin you own. NMC means paying for sheets plus heat paths, and for hard gas control.
3. Treat blast control as life safety. Size the vent area under NFPA 68 or EN 14994 against the gas level at the vent point. Then write crew rules that assume a blast mix before anyone opens a door.
4. Use charge limits as depth, not as proof. Idle units at 30–50% help, though the safe point shifts with heating rate and layout.

For EV and pack engineers

5. Take GB 38031-2025 as the global target if you sell into China or buy from it. The 120-minute rule now sets the bar that most OEMs design to anyway.
6. Stop side-wall tears in dense round cells. A tear defeats aimed venting and hands the fire to the next cell at once, so favour strong cans, bottom vents, or metal spacers.

For fire chiefs, insurers and researchers

7. Read a UL 9540A report as data, never as a badge. Check the trigger method, the charge state, the edition, and the cell age. Big sites merit multi-rack tests.
8. Push for common reporting of step times and charge limits. Published numbers still resist comparison across labs. The gas hazard deserves the same care.

Two shifts would change this advice. Cells with a high enough T2, whether solid-state or sodium-ion, could satisfy the balance with no sheets at all. A common trigger method plus aged-cell testing would also make UL 9540A reports comparable side by side.

Key takeaways

Thermal runaway propagation obeys arithmetic, not luck. One cell gives off heat, and four paths carry a share of it next door. Thermal runaway propagation then runs only while that share tops what the neighbour needs to reach some 200–260 °C. Since about 12% of one cell’s output does the job in a tight module, any real fix must block nearly the whole path.

The design that works looks dull. Aerogel or mica sheets a millimetre or two thick, metal spreaders that drain heat away, vents aimed into a duct, and space between units. Suppression still belongs in the plan, though it earns its keep outside the module, not inside it.

Rules now close on the same target from two sides. China asks for the outcome outright, with no fire and no blast for 120 minutes. North America and Europe still ask for the process. Both roads lead to one balance. Keep the delivered heat below what the next cell needs, prove it at module level, and the box never becomes the story.

Cite this articleHow to cite

Dinh, D. C. (2026, August 21). Thermal Runaway Propagation: How One Cell Takes a Pack. PyroRisk. https://pyrorisk.net/blog/thermal-runaway-propagation-how-one-cell-takes-a-pack/

Get citation →