Process Optimisation and Intelligent Control

We connect predictive models with decisions about process operation. Our research investigates flexible operating regions, optimisation under uncertainty and intelligent control strategies for batch processes and biological production systems.

The challenge

Process performance depends on decisions about operating conditions and how those conditions change over time. A decision that performs well for a nominal model may be unsuitable when raw materials, measurements or process behaviour vary. Quality requirements and operating constraints therefore need to be considered alongside productivity.

We investigate how predictive models, optimisation and intelligent control can support these decisions. Our work includes identifying acceptable operating regions and developing control strategies that account for uncertainty and differences between a model and the process it represents.

Our approach

Identifying flexible operating regions

Operational-space methods identify ranges of operating conditions that can meet process requirements. Our surrogate-enhanced framework uses computationally efficient approximations to help identify flexible operating regions and then locate favourable subregions through dynamic optimisation. This provides alternatives to relying on a single nominal operating point.

Learning control policies under uncertainty

We also investigate reinforcement learning for batch process control. A controller learns a decision policy by interacting with a process model and evaluating the consequences of its actions. Our chance-constrained framework uses Gaussian-process uncertainty estimates to account for constraints and plant–model mismatch, with performance assessed against nonlinear model predictive control in case studies.

Controlling metabolic activity over time

In dynamic metabolic control, operating decisions can extend to changes in enzyme expression during a bioprocess. Published work trains control policies using surrogate models and varies system parameters during training to assess robustness. The demonstrations use simulated fatty-acid and lactate production systems, where productive metabolic activity must be balanced against cellular burden.