Unsupervised Motif Discovery in Engineering Processes: A Machine Learning Approach Using Graph-Based Segment Embeddings

Introduction: The effective reuse of complex simulation models in process engineering requires automated methods to extract domain knowledge from their topological structure. This proof-of-concept study addresses the limitations of holistic, text-reliant model analysis by proposing an unsupervised approach for discovering structural motifs in System Dynamics models.

Methods: To mitigate lexical bias, textual metadata was removed from the dataset. Simulation models were parsed into typed structural segments (Feedback Loops, Causal Paths, and Stock-Flow Subsystems) and transformed into feature vectors representing their mathematical and graph topology. To evaluate the methodology, we used an expanded corpus of 46 engineering models across four domains. Using K-Means clustering (K=4, aligned with the target domain count), we compared our 24-dimensional topological embedding against a whole-model baseline and two intermediate representations (bag-of-nodes and topology-lite).

Results: The full topological configuration achieved an average Adjusted Rand Index (ARI) of 0.2465. While modest in absolute terms for standard clustering tasks, this zero-text performance represents a 6.6-fold improvement over the whole-model baseline (ARI = 0.0374). The intermediate baselines (bag-of-nodes ARI = 0.1244, topology-lite ARI = 0.1145) indicated that increasing structural complexity enhances domain separation. Qualitative analysis showed that chemical processes and bio-engineering models frequently clustered together, reflecting their shared mathematical foundations in mass-action kinetics.

Conclusion: By demonstrating that localized structural motifs can capture domain information more effectively than holistic representations, this study provides a foundation for automated, topology-driven knowledge discovery in systems engineering.