Table of Contents
Reactor optimisation with machine learning means using models learned from experiments or plant data to find the temperatures, pressures, feed ratios and residence times that give the best yield, selectivity or energy use within safety limits. Today it works best in the lab and pilot plant, or as an advisor to existing control, rather than as an autonomous controller.
Key takeaways
- In the lab, Bayesian optimisation drives "self-optimising" reactors. One automated flow platform reached an 81% overall yield in about 14 hours for a telescoped synthesis.
- Plant data from normal operation rarely covers enough of the operating range. In the model predictive control (MPC) plant tests described in an industry survey, each manipulated variable is stepped 8 to 15 times over up to 5 to 15 days.
- Hybrid models keep the mass balances and kinetics and learn only the unknown part. In one industrial batch reactor, a rate constant still had to be refitted 30% lower at plant scale.
- Reinforcement learning is still mostly research for reactors. What Yokogawa describes as the first formal adoption of it for direct plant control was on a distillation column.
- Every optimiser must stay inside the safety envelope: reaction rate roughly doubles with every 10 °C rise.
What does optimising a reactor mean?
A reactor optimisation problem has three parts.
- The objective: conversion, selectivity, yield, throughput, energy use per tonne, catalyst life, or a margin that combines them.
- The decision variables: temperature or temperature profile, pressure, feed ratios, residence time or feed rate, catalyst dosing, batch time.
- The constraints: maximum temperature and pressure, cooling capacity, product specification, and what the downstream separation can handle.
The constraints matter more than people expect. A classic survey of industrial MPC notes that the economic operating point of a typical process unit often lies at the intersection of constraints. Optimising a reactor usually means moving closer to those limits without crossing them, which is why model accuracy near the limits is what counts.
Where does machine learning fit in plant control?
The same survey describes the usual hierarchy: a plant-wide economic optimisation run about once a day, a local unit optimisation about every hour, MPC about every minute, and basic PID control in the DCS every second (Qin and Badgwell, 2003).
Machine learning can enter at each level: as the model inside an optimiser, as the model inside MPC, or as a soft sensor that estimates a quality variable the others need. Most industrial MPC still uses linear models identified from plant tests. In the mid-1999 vendor survey, about 4,500 linear MPC applications were counted against 93 nonlinear ones, and the areas with most nonlinear applications were chemicals, polymers, and air and gas processing.
Which machine learning approaches are used for reactors?
| Approach | Learns from | Where it is used | Main risk |
|---|---|---|---|
| Bayesian optimisation of experiments | Lab or pilot runs chosen one batch at a time | Lab and pilot reactors | Optimum found at small scale may not transfer |
| Surrogate of a process simulator | Runs of a rigorous model | Design and what-if studies | Inherits the simulator's errors |
| Data-driven plant model plus optimiser | Historian and lab data | Advisory setpoints | Correlated, narrow data; extrapolation |
| Hybrid model (kinetics plus learned terms) | Lab and plant data | Scale-up, unit optimisation, MPC | More modelling effort |
| Nonlinear model inside MPC | Plant step tests | Closed-loop control | Test cost and model upkeep |
| Reinforcement learning | Simulator, then plant | Research and first trials | No built-in safety or stability guarantee |
The second row is covered in our guide to surrogate models for chemical processes.
How does Bayesian optimisation work for lab and pilot reactors?
A self-optimising reactor combines three things: a reactor that can run unattended, usually in continuous flow; online analysis; and an algorithm that reads each result and chooses the next conditions.
Clayton and colleagues built such a platform for a telescoped synthesis, using a single online HPLC that sampled at each reactor outlet. The optimiser adjusted four variables, including the residence time and temperature of the first reactor, and identified an 81% overall yield in about 14 hours.
Real reactors rarely have one objective. Schweidtmann and colleagues used a multi-objective machine learning algorithm on two reactions in continuous flow. It found the trade-off curve between environmental and economic objectives in both cases, with fewer experiments than single-objective optimisations needed and without arbitrary weighting factors.
The reactor is also not the whole process. A follow-up study optimised reactions together with downstream work-up, and showed how the optimal conditions are re-evaluated when work-up requirements change. The best conversion is not always the best plant result.
Two limits apply. Bayesian optimisation is best suited to fewer than 20 continuous variables and to experiments that take minutes or hours. And benchmarks against people are encouraging but specific: on a palladium-catalysed reaction, it outperformed expert chemists in average efficiency and consistency. Our comparison of design of experiments and Bayesian optimisation explains when a classical design is still the better choice.
Why is plant data harder to learn from?
On a production reactor, nobody can run free experiments. The historian holds years of data, but controllers keep most variables near their setpoints, inputs move together, and the range covered is narrow. A model can fit that history well and still predict the wrong direction when an optimiser moves one variable alone.
MPC practice shows what it takes to get informative data. In the plant test protocol described in the MPC survey, test signals must excite both the steady-state and dynamic behaviour of the process. Each manipulated variable is stepped 8 to 15 times, with an output signal-to-noise ratio of at least six, and the test may take up to 5 to 15 days. The survey adds that an empirical nonlinear model needs even more test data than a linear one.
The practical answer is to plan small, approved step tests rather than hoping the history contains the answer. Our guide to extracting historian data for machine learning covers the data side.
Why do hybrid models matter for reactors?
A hybrid model keeps what is known, such as mass balances, energy balances and the reaction scheme, and uses a data-driven component only for what is not. A review of hybrid modelling in process engineering concludes that it does not automatically give better results, but that rational knowledge integration can significantly improve model-based operation and design.
An ETH Zurich industrial case shows why. Kinetics fitted on 1 kg lab experiments described the lab data with an R² of about 98%. In the 6 m³ production vessel, removal of a by-product was scale-dependent, so two neural networks were added to estimate it. Even then, one rate constant had to be refitted about 30% lower at plant scale, possibly because catalyst activity differed. After that, the hybrid model matched plant measurements well.
The lesson: a lab-trained model is a starting point, not a plant model. Plan for recalibration on plant data.
Is reinforcement learning ready for reactor control?
Not for most exothermic reactors. A 2025 survey of reinforcement learning in process systems engineering notes that closed-loop stability guarantees remain critical in practice but are not inherently addressed by standard formulations, and that the cost of data acquisition is a significant obstacle in batch processing.
The best-known industrial case is not a reactor. In 2023, Yokogawa announced that its reinforcement learning AI was formally adopted to control a distillation column at an ENEOS Materials plant after a field test of almost a year. The company reports a 40% reduction in the steam previously used to maintain liquid level, compared with manual control. That is a vendor claim about a different unit type, not a reactor benchmark.
What safety limits must an optimiser respect?
The UK Health and Safety Executive explains the core risk. Thermal runaway begins when the heat produced by a reaction exceeds the heat removed. Heat removal rises linearly with temperature, while heat production rises exponentially, and a rule of thumb is that the reaction rate doubles with every 10 °C rise. Scale makes it worse, because heat produced grows with volume while cooling depends on surface area.
The HSE lists the causes of incidents as inadequate understanding of the chemistry and thermochemistry, inadequate heat removal design, inadequate control and safety systems, and inadequate procedures. In December 2007, a runaway reaction destroyed a 2,500-gallon batch reactor at T2 Laboratories in Florida, killing four employees. The US Chemical Safety Board found a likely cooling failure and an unintended second exothermic reaction that occurred above 390 °F, slightly above normal production temperature.
For a machine learning optimiser, this gives firm rules:
- Set its bounds inside the envelope established by calorimetry, HAZOP and the process safety team, never from the data alone.
- Keep trips and safety instrumented systems outside its reach.
- Treat any new operating range as a modification under your management of change procedure.
- Model the constraints, such as cooling duty and peak temperature, not only the objective.
How do you start, step by step?
- Choose one reactor and one objective. Put a value on it, for example the margin from one point of selectivity.
- Write down the constraints with process safety. Maximum temperature, pressure, cooling duty, product specification and downstream limits.
- Audit the data. Historian tags, lab results with sampling times, feed analyses, catalyst age and events such as shutdowns.
- Pick the approach where the uncertainty is. Bayesian optimisation for lab and pilot work, a surrogate when a trusted simulator exists, a hybrid model when kinetics are known but plant behaviour is not.
- Build a simple baseline first. Test it on a later period and check the direction of each effect with process engineers.
- Run in advisory mode. Let the optimiser propose small moves within a narrow band, let operators decide, and log predicted against actual results. Our guide to integrating AI with SCADA and OT systems covers the connection.
- Widen the range with planned step tests, each approved through management of change.
- Monitor drift. Process dynamics change with operating conditions and catalyst ageing. Our guide to model drift monitoring explains what to track.
What we have seen at KABAS AI
Our Virtual Lab is built to test formulations and process parameters virtually before the lab or pilot phase. The two deployments we describe publicly are formulation projects, not reactor control, and we have not published a reactor optimisation project.
Lessons from them still carry over. At ENAD Shymeca, historical formulation data was migrated into the platform before it was used for prediction, and the case study reports a 40% reduction in physical testing cycles. At ADFC (Sonatrach Group), multi-criteria optimisation balances several targets at once, and comparisons between simulated and actual results guide refinements.
For reactors, the same discipline applies: start in the lab or pilot plant, where experiments are cheaper and safer, confirm every prediction against a measurement, and move to the plant in advisory mode first. Our virtual lab for industrial R&D guide explains the approach, the chemistry solutions page shows how it fits a chemical company, and our data science team builds models on industrial data and deploys them on your own infrastructure. For batch reactors, see also golden batch analysis.
Sources
- Qin and Badgwell, A survey of industrial model predictive control technology, Control Engineering Practice 11, 2003
- Clayton et al., Bayesian Self-Optimization for Telescoped Continuous Flow Synthesis, Angewandte Chemie International Edition, 2023
- Schweidtmann et al., Machine learning meets continuous flow chemistry: Automated optimization towards the Pareto front of multiple objectives, Chemical Engineering Journal 352, 2018
- Clayton et al., Automated self-optimisation of multi-step reaction and separation processes using machine learning, Chemical Engineering Journal 384, 2020
- Frazier, A Tutorial on Bayesian Optimization, arXiv 1807.02811, 2018
- Shields et al., Bayesian reaction optimization as a tool for chemical synthesis, Nature 590, 2021
- von Stosch et al., Hybrid semi-parametric modeling in process systems engineering: Past, present and future, Computers and Chemical Engineering 60, 2014
- Simon, Fischer and Hungerbühler, Modeling of a three-phase industrial batch reactor using a hybrid first-principles neural-network model, ESCAPE17, 2007
- Bloor et al., Survey and Tutorial of Reinforcement Learning Methods in Process Systems Engineering, arXiv 2510.24272, 2025
- Yokogawa, Autonomous control AI officially adopted for use at an ENEOS Materials chemical plant, press release, March 2023
- UK Health and Safety Executive, Chemical reaction hazards and the risk of thermal runaway, INDG254
- US Chemical Safety Board, CSB finds T2 Laboratories explosion caused by failure of cooling system resulting in runaway chemical reaction
Frequently asked questions
Can machine learning replace a kinetic model of my reactor?
Usually not, and it rarely needs to. A kinetic model carries mass and energy balances and reaction mechanisms that a pure data model cannot learn from a limited operating range. A hybrid model keeps that structure and uses machine learning only for the part that is poorly known, such as a scale-dependent removal rate or a catalyst activity term.
How many experiments does a self-optimising reactor need?
There is no fixed number. It depends on how many variables are varied, how noisy the analysis is and how the first experiments are chosen. Bayesian optimisation is best suited to fewer than 20 continuous variables, and one published automated flow platform reached an 81% overall yield for a telescoped synthesis in about 14 hours.
Can we optimise a reactor from historian data alone?
Only partly. Under normal operation, controllers keep most variables close to their setpoints, so the history covers a narrow range and inputs tend to move together. A model trained on that data can predict well and still give the wrong direction when an optimiser moves one variable on its own. Small, planned step tests are usually needed to fill the gaps.
Should a machine learning optimiser write setpoints directly to the DCS?
Not at first. Start in advisory mode, where the optimiser proposes moves within a narrow band and operators decide, and compare predicted with actual results. Closed-loop use comes later, through the existing advanced control layer and within limits set by process safety. Trips and safety instrumented systems stay outside the optimiser's reach.
Is reinforcement learning used to control reactors in industry?
Rarely. Yokogawa describes its 2023 adoption at an ENEOS Materials plant as the first formal use of reinforcement learning for direct control of a plant, and the unit was a distillation column, not a reactor. Surveys of the field still point to data cost and the lack of built-in stability guarantees as obstacles.
Can these models run on our own servers?
Yes. Surrogate, hybrid and Bayesian optimisation models are small and need no internet connection once trained. Because they encode reaction conditions and plant know-how, many companies keep both the data and the models on their own infrastructure.



