Text-to-Flowsheet: An Automated LLM-Pipeline for Digitization and Simulation of Chemical Processes with Expert-Level Accuracy

Converting unstructured natural language descriptions into structured process flowsheets is a fundamental bottleneck in chemical engineering, traditionally requiring years of expert training. While large language models (LLMs) show promise in text comprehension, their ability to match human expertise in modeling complex chemical process flowsheets remains unproven. Here, we present a rigorous benchmark comparing a fully automated LLM-powered digitization pipeline against the collective performance of 50 chemical engineering experts. Our pipeline leverages LLMs to extract process structures from text and formalize them as flowsheet graphs. To handle the inherent ambiguities of natural language, we utilize constrained, step-by-step prompting augmented with thermodynamic property calculations. Subsequently, the digitized flowsheet graphs are automatically translated into the flowsheeting software Aspen Plus to compute rigorous mass and energy balances. Black-box optimization on subprocess structures is used to estimate unknown parameters and ensure simulation convergence, completing the pipeline from text to converged process simulation. For the first time, we demonstrate that an automated pipeline can achieve expert-level accuracy in process digitization. Using a unique, newly-generated dataset of 101 expert-drawn flowsheets, we show that our LLM-assisted approach faithfully captures process topology and operating conditions even in the face of incomplete information. This work provides a robust, validated framework for the large-scale digitization of chemical production literature, contributing a transformative tool and dataset for the community to accelerate automated process design and assessment.