The workflow addresses the challenge of understanding how dynamic patterns in large spatiotemporal ensembles vary with experimental input parameters. In the motivating domain of droplet impact experiments, researchers seek to map how splash phenomena (e.g., crown formation, jetting, bubble-splash) relate to physical parameters such as Weber number and dimensionless film thickness—a process that traditionally requires hundreds of hours of manual video inspection.
Preparation. A self-supervised Siamese neural network extended with a spatiotemporal adaptive pooling layer is trained on the ensemble data, learning a similarity metric that accommodates variable-sized spatiotemporal patches. The ensemble is then uniformly sampled to produce a precomputed search space of patches paired with their latent-space embeddings. This search space is clustered (K-means) and projected (UMAP) to create an overview arrangement of pattern similarity structure.
Initial visualisation. Two coordinated views are generated before the exploration loop: a direct data visualisation allowing the analyst to browse any ensemble member's frames and draw bounding boxes for patch selection, and a cluster overview displaying flower glyphs at UMAP-projected centroid positions (initially with empty petals since no queries exist yet).
Iterative pattern discovery (loop). The analyst inspects the cluster overview to identify clusters with no or few query matches (uncolored glyphs), navigates to their contents or to ensemble member data, and recognises a distinct spatiotemporal pattern (e.g., "crown splash"). The analyst then defines a query by selecting positive and negative example patches and adjusting a match-score threshold via an interactive score graph. The system scores all search-space samples against the full set of active queries and determines matches. Two coordinated visualisations update: timeline views (superimposed line charts per member showing temporal match distributions) and the cluster overview (flower glyphs now showing match proportions per query with density contours). The analyst assesses pattern coverage by checking whether uncolored clusters remain. If coverage is insufficient, the loop continues with another query; otherwise, the analyst exits.
Post-loop analysis. Temporal pattern sequences are computed for each ensemble member by binning match counts over time intervals and assigning the dominant query per interval. Members sharing identical sequences are grouped. A sequence explorer lists these groups, and a linked parameter-space scatterplot visualises where each group's members fall in parameter space. The analyst interactively selects groups to highlight in the parameter space, identifies spatial clustering of groups within parameter regions, and derives a characteristic map describing which parameter configurations produce which pattern dynamics.