Abstract for: A Tree-Based Multi-Output Clustering Approach for Pattern-Oriented Analysis of System Dynamics Model Outputs

System dynamics models exhibit a wide range of behavioral patterns depending on their parameter settings. However, systematically connecting these patterns to the underlying parameters that generate them remains a challenge. Existing approaches often require predefined behavior labels and expert classification, or separate clustering and rule extraction into disconnected, two-step processes This study applies a tree-based clustering approach that groups similar generated trajectories and extracts interpretable if-then rules within a single structure. The method combines Latin Hypercube Sampling for parameter exploration, Dynamic Time Warping for similarity measurement, and recursive tree partitioning. The procedure derives clusters directly from the simulation trajectories without assuming prior behavior labels. The approach was applied to two models using 1000 simulation runs each. In the temperature adjustment model, eight clusters emerged, ranging from strongly oscillatory adjustment to smooth goal-seeking convergence. In the World3-03 model, fifteen trajectory groups were produced, each defined by explicit parameter thresholds, allowing the analyst to survey the full behavioral range. The findings demonstrate that clusters and their parameter thresholds can emerge without predefined behavior labels or expert classification. Since the tree-based structure groups trajectories and produces rules simultaneously, the disconnected two-step process of previous work is avoided. This partitioning of the simulation ensemble provides an interpretable mapping between parameters and the resulting system response. grammatical refining