Reliable process control, tested against evidence.

Research question

In a dynamic chemical-process simulation, can a reinforcement-learning controller improve multi-objective operation while satisfying temperature, flow and other operating constraints under disturbances?

Experimental design

• Define a continuous stirred-tank reactor (CSTR) model with documented assumptions, state variables, manipulated variables and operating limits.
• Implement conventional baselines, including proportional-integral-derivative (PID) control and, where appropriate, model-predictive control (MPC).
• Define training and evaluation environments separately to reduce the risk of reporting training performance as generalisation.
• Specify reward terms and independent safety/constraint checks before evaluating policies.
• Test set-point tracking, settling behaviour, overshoot, disturbance rejection, constraint violations, energy/resource proxies and robustness across parameter shifts.
• Record seeds, software versions, parameter sets, evaluation protocols and known limitations.

Reporting policy

Performance improvements, statistical comparisons and claims of industrial relevance will be reported only after measurements exist and the experimental procedure can be inspected. This page does not present projected targets as observed results.

Safety boundary

Results from a simulator are not evidence that a controller is safe for a real plant. Physical deployment would require validated process models, independent engineering assessment, fail-safe controls, appropriate approvals and authorised site access.