An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models

Research Approach

This paper introduces an automated, agentic workflow designed to bridge the gap between dynamic process modeling and industrial control implementation. Rather than relying on manual engineering, the framework uses a large language model (LLM) to decompose the control design process into a sequence of constrained, executable code-generation tasks. These tasks include plant normalization, manipulated-variable/controlled-variable (MV-CV) pairing, controller specification, and the construction of a closed-loop simulation environment.

Workflow and Validation

A critical feature of the proposed system is its iterative validation and repair loop. Each generated code artifact is subjected to automated checks for syntax, imports, and interface consistency. If a module fails validation, the system provides the error feedback to the LLM to trigger a repair. This ensures that only functional, interface-compliant code proceeds to downstream stages. The final stage of the pipeline uses Bayesian optimization to tune the controller parameters against a deterministic evaluation scenario, which is itself automatically generated based on the process dynamics.

Performance and Results

The framework was tested on a nonlinear gas-preheater benchmark characterized by coupled pressure and temperature dynamics. The LLM successfully generated a physically consistent decentralized PI feedback-feedforward control structure. Following the automated tuning phase, the closed-loop performance objective—which aggregates set-point tracking and disturbance-rejection errors—was reduced by approximately 26.5% compared to the initial controller generated by the workflow. This result confirms the feasibility of using LLM-based code generation to construct end-to-end control-design pipelines.