Predictive Thermal Runaway

The system ingests a synchronized stream of voltage, current, internal temperature, surface temperature, pressure, and off-gas volatile organic compound (VOC) signatures. Unlike simple threshold monitoring, sensor fusion algorithms cross-correlate these disparate signals to detect subtle precursor patterns—such as a decoupling of internal and surface temperature rise rates—that indicate the onset of exothermic decomposition at the solid-electrolyte interphase (SEI) layer.

At the core of prediction lies the Arrhenius equation, which quantifies the exponential relationship between temperature and the rate of internal chemical decomposition reactions:

• Models the SEI decomposition stage (~80-120°C)
• Tracks the anode-electrolyte reaction onset (~120-150°C)
• Forecasts the critical separator meltdown threshold (~160-180°C)
• Calculates the self-heating rate (°C/min) to project time-to-catastrophe

This physics-informed approach allows the system to extrapolate beyond observed data points and estimate the remaining safe operating window in seconds.

Before a cell enters full thermal runaway, it emits a characteristic sequence of electrolyte vaporization gases. The system employs metal-oxide semiconductor (MOS) and non-dispersive infrared (NDIR) sensors to detect:

• Carbon monoxide (CO) — early indicator of SEI breakdown
• Hydrogen (H₂) — signals lithium plating and electrolyte reduction
• Hydrogen fluoride (HF) — indicates separator decomposition
• Methane (CH₄) and ethylene (C₂H₄) — late-stage venting precursors

A rising concentration gradient across these gas species, analyzed via a trained classification model, provides a chemical fingerprint of imminent failure distinct from normal operational off-gassing.

Prediction latency is safety-critical. The inference engine runs directly on an edge gateway or battery management system (BMS) microcontroller using TinyML optimized models:

• Quantized neural networks compressed to run on ARM Cortex-M class processors
• Sub-millisecond inference latency eliminates cloud round-trip delays
• Local processing ensures functionality during connectivity loss
• Federated learning updates improve model accuracy across fleets without exposing proprietary cell performance data

This architecture guarantees that a thermal runaway warning is issued within the narrow window required for automated load shedding or fire suppression activation.

In multi-cell battery packs, a single cell entering runaway can trigger a cascading propagation event. The system models:

• Inter-cell thermal conductance based on pack geometry and cooling system design
• Radiative and convective heat transfer to adjacent cells
• Ejecta and hot particle dispersion patterns from venting cells
• State of charge (SOC) of neighboring cells as a risk multiplier

This propagation model predicts not just the first cell failure, but the spatial and temporal sequence of subsequent failures, enabling targeted suppression at the source before the cascade becomes unstoppable.

Prediction without action is insufficient. Upon crossing a defined probability threshold (typically >95% confidence of imminent runaway within 30-120 seconds), the system autonomously initiates:

• Contactor opening to electrically isolate the affected module
• Maximum cooling system activation to extract heat from the pack
• Fire suppression agent release in stationary storage applications
• Emergency vehicle shutdown and occupant alert in automotive contexts
• Telemetry broadcast to facility safety systems and first responders

These actions execute via deterministic, hard-coded logic paths that bypass higher-level software stacks to eliminate latency from operating system scheduling or network congestion.

The foundational kinetic model that quantifies how the rate of a chemical reaction increases exponentially with temperature. In thermal runaway prediction, it models the self-accelerating decomposition of electrolytes or active materials.

• Activation Energy (Ea): The energy barrier that must be overcome for a reaction to proceed
• Frequency Factor (A): The pre-exponential constant representing molecular collision frequency
• Temperature Dependence: A 10°C rise typically doubles the reaction rate in lithium-ion systems

Real-time sensor data feeds into Arrhenius-based models to calculate the heat generation rate and forecast when it will exceed the system's heat dissipation capacity—the critical tipping point into runaway.

The automated collection and wireless transmission of real-time environmental data from monitoring devices to a central analytics platform. For thermal runaway prediction, this includes:

• Temperature sensors: Thermocouples and RTDs capturing internal cell temperature at sub-second intervals
• Voltage sensors: Monitoring individual cell voltage for anomalous drops indicating internal short circuits
• Gas sensors: Detecting off-gassing of volatile organic compounds (VOCs) like ethylene carbonate before thermal runaway
• Pressure sensors: Identifying cell swelling from gas buildup within sealed enclosures

Sensor fusion across these modalities provides the multi-dimensional data stream required for predictive algorithms to distinguish normal operation from pre-runaway signatures.

The execution of trained machine learning models directly on local hardware—such as a Battery Management System (BMS) microcontroller—without requiring cloud connectivity. This is critical for thermal runaway prediction because:

• Latency elimination: Detection decisions must occur in milliseconds; round-trip cloud latency is unacceptable
• Offline resilience: Battery systems in remote locations or underground cannot rely on network availability
• Data privacy: Raw telemetry never leaves the device, satisfying security requirements

TinyML models compressed via quantization and pruning run anomaly detection algorithms directly on the BMS, issuing immediate shutdown commands when a runaway precursor pattern is identified.

A dynamic, virtual replica of a physical battery pack or chemical storage system that ingests real-time sensor data to simulate thermal behavior. In predictive thermal runaway contexts, the digital twin:

• Runs parallel simulations of heat generation and dissipation using finite element analysis
• Compares predicted vs. actual temperature curves to detect deviations indicating abnormal self-heating
• Stress-tests failure scenarios by simulating internal short circuits or coolant system failures
• Updates continuously as the physical asset ages, incorporating State of Health (SoH) degradation

The twin provides a physics-informed baseline against which machine learning anomaly detectors can calibrate their sensitivity, reducing false positives.

A statistical methodology that moves beyond correlation to identify the specific root cause of a thermal event. While predictive models flag that runaway is imminent, causal inference answers why it is occurring:

• Directed Acyclic Graphs (DAGs): Model causal relationships between variables like charge rate, ambient temperature, and internal resistance
• Counterfactual analysis: Determines whether a different charging protocol would have prevented the event
• Do-calculus: Isolates the effect of specific interventions, such as reducing charge current

This enables Corrective and Preventive Action (CAPA) —not just stopping one runaway event, but redesigning operational parameters to eliminate the causal pathway entirely.

A calculated single temperature value that simulates the cumulative thermal stress on a substance over time, weighted using the Arrhenius equation. While traditionally used in pharmaceutical cold chains, MKT is adapted for battery health monitoring:

• Cumulative degradation metric: Integrates all temperature fluctuations into one equivalent constant-temperature exposure
• Non-linear weighting: Higher temperatures contribute disproportionately to the MKT value
• Predictive input: Rising MKT trends serve as a leading indicator of accelerated SEI (Solid Electrolyte Interphase) growth, a precursor to internal shorts

Tracking MKT alongside instantaneous temperature gives predictive models a long-term thermal history context that improves runaway forecasting accuracy.