Sketch2Simulation: Automating Flowsheet Generation via Multi Agent Large Language Models

A team of researchers has introduced Sketch2Simulation, a novel framework that uses coordinated multi-agent large language models (LLMs) to automate a critical bottleneck in process systems engineering. The system takes a hand-drawn or digital sketch of a chemical process flowsheet and translates it directly into a fully executable simulation model for the industry-standard Aspen HYSYS software. It works through three specialized layers: a diagram parsing layer that interprets visual elements, a synthesis layer that constructs a graph-based intermediate representation and generates code for the HYSYS COM interface, and a validation layer that performs multi-level checks on the generated model.

In evaluation across four chemical engineering case studies of increasing complexity—from a simple desalting process to an industrial aromatic production flowsheet with multiple recycle loops—the system successfully produced executable HYSYS models in all cases. It demonstrated high accuracy, achieving connection consistency above 0.93 and stream consistency above 0.96. The research, detailed in a new arXiv paper, represents a significant step toward end-to-end automation in process design, though challenges remain in handling dense recycle structures and implicit diagram semantics. This work bridges the gap between previous AI tools that only extracted graphs from diagrams and those that required structured textual inputs, creating a true sketch-to-simulation pipeline.