LLM Agents Achieve 94% Success Rate in PID Tuning for Chemicals

A team from Zhejiang University has introduced a novel framework that leverages language model agents for PID tuning in chemical processes. Traditional PID tuning relies on identified process models or manual iterative adjustments by engineers. The new framework formalizes this engineer-like workflow: the LLM receives closed-loop response features, control-engineering diagnoses, and tuning preferences, then generates and iteratively corrects PID gains using internal model control (IMC) demonstrations. It works with both large hosted models (e.g., DeepSeek-V4-Flash, Qwen3.7-Plus) and smaller local models (Qwen3-0.6B) fine-tuned via supervised fine-tuning (SFT) and physics-informed group relative policy optimization (PI-GRPO). The PI-GRPO method incorporates non-compensable stability and performance rewards to ensure robust gains.

Results on 100 first-order plus dead time (FOPDT) and 100 second-order plus dead time (SOPDT) test cases show hosted LLMs achieved final success rates of 75-89% (DeepSeek-V4-Flash) and 77-79% (Qwen3.7-Plus). For the smaller Qwen3-0.6B model, SFT raised first-recommendation success to 86.5%, and PI-GRPO further boosted it to 94.0%—primarily improving first-attempt reliability and stability margins. This work demonstrates that even small language models, when properly trained with physics-informed rewards, can rival or exceed the performance of much larger hosted models for specialized control engineering tasks.