Thermal Runaway Modelling at Cell Level

The physics of thermal runaway describes the chain of exothermic reactions inside a cell once it is triggered. Modelling that chain at cell level means turning those reactions into equations that can predict, ahead of a physical test, roughly when a cell will go into runaway and how quickly it will self-heat once it does. This is a different problem from modelling how runaway propagates between cells once one cell has already failed – that is a pack-level, cell-to-cell heat transfer problem covered separately. Cell-level models instead treat a single cell (or a small representative volume of its jellyroll) as the unit of analysis, and are the building block that pack-level propagation models and BMS safety functions both rely on for their inputs.

Lumped Kinetic (Semi-Empirical) Models

The most widely used family of cell-level models treats the whole jellyroll as a single lumped thermal mass and represents thermal runaway as a small number of parallel decomposition reactions – typically SEI breakdown, anode-electrolyte reaction, cathode decomposition and electrolyte decomposition – each following Arrhenius-type kinetics. The heat generated by reaction i is written as a rate term multiplied by a reaction enthalpy:

\dot{q}_i = m_i \, \Delta H_i \, A_i \, e^{-E_{a,i}/RT} \, f(c_i)

where m_i is the mass of reactive material, ΔH_i the reaction enthalpy, A_i and E_a,i the Arrhenius pre-exponential factor and activation energy for reaction i, and f(c_i) a function describing how the remaining reactant concentration limits the rate. Summing q̇_i across the reaction set and coupling it to a lumped heat-capacity balance gives a self-heating curve that can be compared directly against test data.

This approach was established by Hatchard et al., who fitted a four-reaction lumped model (SEI, anode, cathode and electrolyte decomposition) to accelerating rate calorimetry (ARC) data from 18650 LCO cells and used it to reproduce hot-box abuse test results (Hatchard, MacNeil, Basu & Dahn, 2001). Most subsequent cell-level models are extensions of this same structure – adding reactions, moving to other chemistries such as NMC and LFP, or resolving the jellyroll into multiple radial or axial lumps instead of one – rather than a fundamentally different approach.

The Role of ARC and DSC Data

Lumped kinetic models are only as good as the kinetic parameters (A_i, E_a,i, ΔH_i) fitted into them, and those parameters are normally extracted from accelerating rate calorimetry (ARC) and differential scanning calorimetry (DSC) measurements. ARC in particular is run adiabatically, so the self-heating rate it records as a function of temperature is close to what the cell would actually do in an insulated fault condition, making it the standard source of onset-temperature and self-heating-rate data that a lumped model is fitted against (Zhang et al., 2026).

Coupled Electrochemical-Thermal Models

A higher-fidelity alternative couples a detailed electrochemical model – typically a Doyle-Fuller-Newman or Single Particle Model style formulation – to the same abuse-reaction kinetics used in the lumped approach, so that internal state (local SOC, current distribution, internal short-circuit resistance) feeds directly into where and how fast heat is generated. This captures effects a purely lumped model cannot, such as how an internal short forms and grows, at the cost of much greater computational effort, and is generally reserved for detailed design verification work rather than routine screening.

Where This Fits

A cell-level model’s output – onset temperature, self-heating rate, peak temperature and total heat release – is the input that pack-level propagation models and pack modelling techniques then use to predict whether and how fast a fault spreads to neighbouring cells. Getting the cell-level kinetics right first is what makes those larger-scale results meaningful rather than just plausible-looking.