Thermal Runaway Prediction in Batch Reactors

Thermal runaway in a batch reactor is not a sudden event. It begins as a subtle imbalance between heat generation and heat removal that compounds over time. By the time an operator reads an alarm on the HMI, the system may already be past the point where intervention is effective. The question for any prediction system is not whether you can detect the event, but whether you can detect the trajectory that leads to it, early enough to act.

A world model approach to this problem is fundamentally different from a threshold alarm. Threshold alarms tell you when a measurement crosses a line. A world model tells you where the current trajectory is headed, and that answer is available before any threshold has been breached.

The Physics of Batch Reactor Runaway

In a stirred batch reactor running an exothermic reaction, the reaction rate typically follows Arrhenius kinetics: rate increases exponentially with temperature. Heat generation is proportional to reaction rate. Heat removal is proportional to the temperature difference between the jacket and the reactor contents, and to the jacket flow rate. Stable operation requires that the heat removal capacity keep pace with heat generation as temperature rises.

Runaway initiates when heat generation begins to outpace removal. This can happen through several routes: a control valve on the cooling jacket fails to open, the jacket coolant supply temperature rises unexpectedly, the reaction rate is higher than expected due to a feed composition deviation, or the agitator speed drops and reduces heat transfer coefficient. None of these causes a single sharp alarm trigger immediately. What they cause is a change in the rate of temperature rise, which is the signal that precedes runaway.

The critical diagnostic quantity is the temperature rise rate (dT/dt) relative to the expected rise rate given the current control actions and process state. If the observed rise rate is higher than what the world model predicts it should be given the current cooling setpoints and feed rates, something is wrong. The question is whether that discrepancy is transient and self-correcting, or whether it indicates a developing heat balance failure.

What the World Model Learns

A world model for a batch reactor is trained on historical batch runs: the full temperature trajectory, the cooling jacket setpoints, the feed rate profile, and the measured heat removal proxy (often the jacket outlet temperature minus inlet temperature times flow rate). Over many batches, the model learns what the normal temperature trajectory looks like given a specific set of control inputs, and it learns the variability around that trajectory, which represents natural batch-to-batch differences in feed composition, catalyst activity, and equipment condition.

Specifically, the model learns the conditional distribution P(T(t+h) | T(t), u(t), s(t)), where T is reactor temperature, u is the control input vector (cooling setpoint, feed rate, agitator speed), s is any available state proxy (jacket delta-T, pressure, pH if measured), and h is the prediction horizon. For a runaway precursor scenario, we evaluate this conditional distribution at horizons of 5 to 20 minutes: given where we are now and the control actions currently in place, what temperature should we expect in 15 minutes?

The prediction includes uncertainty bands calibrated on historical batch data. Normal operating variance sets the width of the expected band. When current conditions suggest a departure from that band, the system generates an early warning. The warning is not "the temperature is too high" but "the temperature trajectory is heading toward a region that falls outside the model's prediction under current control actions, with the gap widening at a rate consistent with an early heat balance failure."

Where This Differs from Classical MSPC

Multivariate statistical process control (MSPC) based on principal component analysis is the classical approach to batch monitoring. BatchPCA and related methods project high-dimensional batch data onto a low-dimensional subspace and flag deviations from normal operating subspace. This works well for detecting that something is wrong. It is less good at answering what will happen next.

The world model approach is predictive rather than diagnostic. Instead of asking "is the current state unusual relative to historical batches?", it asks "given the current state and control inputs, what state is the reactor heading toward?" For thermal runaway specifically, the relevant question is the future state, not the current state. An MSPC alarm that fires when dT/dt first exceeds historical norms may already be too late for some runaway scenarios. A world model that projects the temperature trajectory forward and compares it against the safe operating region gives the operator more time.

We are not saying MSPC is the wrong tool. It is a good tool for detecting process faults and equipment anomalies. We are saying that for runaway prediction specifically, the extra step of projecting the future trajectory under current control conditions is what makes the prediction actionable early enough to matter.

Practical Considerations for Deployment

Batch-to-batch variability is the main challenge for this application. Unlike a continuous process where steady-state is the baseline, batch reactors have a designed temperature profile that changes across the batch duration. The model must learn what normal looks like at each phase of the batch, not just at a single operating point. This requires segmenting the historical batch runs by phase and calibrating uncertainty bounds separately for each phase, or using a model architecture that explicitly represents batch phase as a conditioning variable.

Feed composition variation is another source of difficulty. If the feed comes from multiple suppliers or has variable composition between deliveries, the reaction kinetics will vary and the normal temperature trajectory will shift. A model that cannot account for feed composition variation will generate false positives on batches with unusual feed. Where inline or at-line feed composition measurement is available, it should be included in the model's conditioning vector. Where it is not, the model's uncertainty bounds in the early batch phase should be widened to reflect this unobserved variability.

What an Early Warning Actually Changes

The value of a 10-to-15-minute thermal runaway warning depends entirely on what the operator can do with that time. In a jacketed batch reactor, the primary response options are: increase cooling jacket flow, reduce feed rate or stop feed addition, begin emergency quench addition if available, or initiate a controlled dump to a vent tank. All of these actions are more effective early than late, and some (emergency dump) are only viable if they can be initiated before temperatures reach the point where the dump system itself could trigger secondary hazards.

This is why the prediction window matters. A 3-minute warning after the first threshold alarm gives less response time than a 15-minute trajectory-based warning before any threshold has been breached. The earlier warning does not guarantee a better outcome, but it extends the option space available to the operator and to the safety interlock system. That is the practical value of predictive monitoring for this application.