LLM-guided Chemical Process Optimization with a Multi-Agent Approach

TL;DR

This work tackles constraint-deficient chemical process optimization by introducing an AutoGen-based multi-agent LLM framework that autonomously infers operating bounds from minimal process descriptions and collaboratively guides optimization using IDAES simulations. Evaluated on the hydrodealkylation process, the approach achieves competitive performance relative to IPOPT and grid search while delivering dramatic wall-time reductions (converging in under 20 minutes). The framework relies on reasoning-enabled LLMs (o3/o1) to perform constraint generation, memory-guided parameter exploration, and interpretable decision making, demonstrating robust convergence where standard models may fail. This strategy lowers the expertise barrier for process optimization, enabling rapid, interpretable optimization for emerging or retrofit processes, and opens pathways for integrating database-driven constraints and hybrid optimization schemes.

Abstract

Chemical process optimization maximizes production efficiency and economic performance, but optimization algorithms, including gradient-based solvers, numerical methods, and parameter grid searches, become impractical when operating constraints are ill-defined or unavailable. We present a multi-agent LLM framework that autonomously infers operating constraints from minimal process descriptions, then collaboratively guides optimization. Our AutoGen-based framework employs OpenAI's o3 model with specialized agents for constraint generation, parameter validation, simulation, and optimization guidance. Through autonomous constraint generation and iterative multi-agent optimization, the framework eliminates the need for predefined operational bounds. Validated on hydrodealkylation across cost, yield, and yield-to-cost ratio metrics, the framework achieved competitive performance with conventional methods while reducing wall-time 31-fold relative to grid search, converging in under 20 minutes. The reasoning-guided search demonstrates sophisticated process understanding, correctly identifying utility trade-offs and applying domain-informed heuristics. Unlike conventional methods requiring predefined constraints, our approach uniquely combines autonomous constraint generation with interpretable parameter exploration. Model comparison reveals reasoning-capable architectures (o3, o1) are essential for successful optimization, while standard models fail to converge. This approach is particularly valuable for emerging processes and retrofit applications where operational constraints are poorly characterized or unavailable.

Figures (4)

• Figure 1: Multi-agent workflow of the LLM-based optimization framework. The ContextAgent generates constraints and detailed process overview from basic process info using domain knowledge, while specialized agents within the GroupChat environment handle parameter introduction, validation, optimization suggestions, and IDAES-based simulation evaluation through iterative refinement until optimal solutions are achieved.
• Figure 2: Visual illustration of the workflow of the SuggestionAgent. The agent receives three key inputs: (1) a structured prompt with optimization instructions, (2) process overview including valid constraint ranges, and (3) evaluation history of previous attempts. The SuggestionAgent analyzes these inputs to generate feasible parameter suggestions that satisfy all constraints, which are then validated and refined through iterative multi-agent interactions.
• Figure 3: Flowsheet of the hydrodealkylation (HDA) process used for IDAES simulation and optimization study. The process consists of mixer M101, heater H101, reactor R101, flash separators F101 and F102, splitter S101, and compressor C101. Feed streams include hydrogen (0.30 mol/s), methane (0.02 mol/s) and toluene (0.30 mol/s) at 303.15 K and 350,000 Pa. The reactor operates at 350,000 Pa with 75% toluene conversion, followed by separation units that produce benzene and toluene product streams, with 20% of the overhead stream purged through splitter S101.
• Figure 4: Convergence comparison across different LLM architectures for cost minimization. Solid dots indicate where models terminated, with dashed lines extending to show their final achieved cost for comparison. Red dashed line represents grid search's optimal cost solution of $$5.660 \times 10^5/yr$. Reasoning models (o3 and o1) successfully converge, with o3 requiring 11 iterations and o1 requiring 14 iterations. Standard models (GPT-4o and GPT-4.1) terminate prematurely after 4 to 5 iterations without effective learning.