CRAFTS: AI Framework for Chemical Process Simulation via LLM Agents

The introduction of CRAFTS (Collaborative Role-Adaptive Fine-Tuning of LLM Agents) marks a significant advancement in automating the creation of executable chemical-process models, a task that has typically required manual input from chemical engineers. As outlined in a paper on arXiv (2608.01369), this framework breaks down the simulation development into defined subtasks, each handled by one of seven specialized roles, with deterministic IDAES/Pyomo gates facilitating transitions. By processing natural-language requests, process flowsheet diagram (PFD) data, and curated chemical-engineering insights, the system's Input Understanding and Intent components extract requirements and constraints. Specialists in visual, topology, and specification convert these into typed simulator contracts, while Debug and Optimization roles assist with bounded repairs and optimization tasks. This approach emulates the structured workflow of chemical engineers, addressing complex decisions related to unit operations, thermodynamics, and optimization. Fine-tuning enhances the performance of three schema-related roles, representing a crucial step in leveraging large language models in specialized engineering fields, potentially minimizing manual labor and reducing error risks in chemical process simulation.