Graph-to-SFILES: Control structure prediction from process topologies using generative artificial intelligence

Researchers developed Graph-to-SFILES, a generative AI model that predicts control structures for chemical process designs from flowsheet topologies using graph neural networks. The model achieves 73.2% top-5 accuracy on 10,000 flowsheets and significantly outperforms sequence-based approaches in small-data scenarios, though performance reverses on larger datasets.

The Graph-to-SFILES model represents an advancement in applying neural architectures to engineering design automation, specifically addressing the control structure prediction phase of P&ID (Piping and Instrumentation Diagram) development. This work demonstrates how graph-based representations can capture the structural relationships in chemical processes more effectively than sequential encodings, leveraging the inherent permutation invariance of graph neural networks.

The research emerged from recognition that previous generative AI applications in chemical engineering relied heavily on sequence-based models, which struggle to represent the interconnected nature of process topologies. By reformulating the problem as graph-to-sequence prediction, the authors created a system where the input naturally mirrors how engineers conceptualize process flows, potentially reducing translation overhead between human design intent and machine learning representation.

The performance trade-off between small-data and large-data regimes carries practical significance for industrial adoption. Organizations with smaller, specialized datasets gain meaningful advantages from the graph-based approach—improving accuracy from 0.9% to 28.4% on 1,000 flowsheets. However, the superiority of sequence models at scale suggests that as companies accumulate larger design datasets, the marginal benefit of sophisticated graph architectures diminishes, following established patterns in deep learning.

This research highlights the broader challenge of domain-specific AI: solutions optimized for one data regime often underperform in another. The authors acknowledge that industry validation remains pending, indicating the gap between academic benchmarks and production requirements. Future development should focus on hybrid architectures that retain graph benefits while scaling efficiency, and on testing the model's robustness across diverse industrial processes beyond the controlled training environment.

• →Graph-based neural networks outperform sequence models on small datasets, improving control structure prediction accuracy from 0.9% to 28.4% with only 1,000 training samples.
• →The proposed GNN architecture achieves 73.2% top-5 accuracy on 10,000 flowsheets, demonstrating viable automation for a previously manual engineering step.
• →Sequence-based approaches reverse their disadvantage at scale, performing better than graph methods on 100,000-flowsheet datasets, indicating architecture selection depends on data availability.
• →SFILES 2.0 notation enables representation of control-extended flowsheets as sequential outputs, creating a practical bridge between graph inputs and engineering-standard outputs.
• →Real-world industrial validation remains incomplete, and effectiveness on diverse production cases at scale has not yet been demonstrated.