New paper: Graph-to-SFILES predicts control structures from process topologies

Our paper "Graph-to-SFILES: Control structure prediction from process topologies using generative artificial intelligence" by Lukas Schulze Balhorn, Kevin Degens and Artur M. Schweidtmann was published in Computers & Chemical Engineering.

Control structure design is an important but time-consuming step in the development of Piping and Instrumentation Diagrams (P&IDs). The Graph-to-SFILES model takes a flowsheet topology as a graph and generates the control-extended flowsheet as a sequence in the SFILES 2.0 notation. The paper compares four graph encoder architectures, including a graph neural network proposed in this work, which performed best.

Trained on 10,000 flowsheet topologies, the model reaches a top-5 accuracy of 73.2%. For a small training set of 1,000 flowsheets, the graph-based model improves the top-5 accuracy from 0.9% to 28.4% compared with a purely sequence-based approach. On a large dataset of 100,000 flowsheets, the sequence-based approach performs better. The results show the potential of graph-based generative AI to accelerate P&ID development when little data is available.

The work is part of our research on autocompletion of engineering diagrams.

Paper: L. Schulze Balhorn, K. Degens, A. M. Schweidtmann (2025). Graph-to-SFILES: Control structure prediction from process topologies using generative artificial intelligence. Computers & Chemical Engineering, 199, 109121. https://doi.org/10.1016/j.compchemeng.2025.109121