ParaGlide supports model developers in interactively exploring and partitioning the parameter spaces of computational simulation models. The central goal is to discover regions in the input parameter space that produce qualitatively different output behaviors, thereby supporting sensitivity analysis, context-dependent optimization, and model comparison. The workflow tightly couples data generation (sampling and simulation) with visual exploration (scatter plot matrices and similarity embeddings) and manual output clustering.
The input is a computational model (simulator) that maps n input parameters to complex outputs (images, time curves, plots). The model developer integrates their simulator code via a compute node interface and specifies an initial Region of Interest (ROI) — a hyperbox-shaped region defined by parameter value ranges — within which to explore. They select a sampling strategy (simple random, lattice, Latin hypercube, or low-discrepancy) and a sample count, balancing computational cost against coverage.
Parameter configurations are generated within the ROI using the selected sampling method. The simulator is executed for each configuration, producing outputs (e.g., spatiotemporal pattern images for biological aggregation, segmented brain images, or fuel cell current density plots). This step may be computationally expensive (minutes to hours per run), motivating iterative coarse-to-fine strategies.
From the raw simulator outputs, the user computes derived variables — features capturing meaningful aspects of output behavior. These may include domain-specific quality measures (Dice coefficient for segmentation accuracy), stability indicators (linear stability analysis for PDE models), performance objectives, or pixel/curve values. Derived variables can also serve as cheap screening proxies to guide expensive simulation only to interesting regions.
To compare outputs globally, the user constructs a similarity measure from derived variable vectors (typically cosine similarity / Pearson correlation). A spectral embedding maps the similarity matrix to 2D positions, placing outputs with similar behavior close together. The embedding is displayed as a scatter plot alongside the scatter plot matrix (SPloM) of all input and output dimensions, enabling linked inspection.
The user inspects the SPloM for an overview of how parameter configurations distribute across dimensions, verifying sampling coverage and identifying dense/sparse areas. They click individual experimental points to inspect detailed outputs (images, plots), transitioning seamlessly between overview and detail. The similarity embedding reveals the global structure of output behaviors — clusters of similar outputs, outliers, and gradients of change.
The user enters an iterative cycle of manual clustering, assessment, and refinement. In the embedding view, they draw rectangular selections around groups of similar outputs and assign cluster labels (colour-coded). They verify cluster coherence by inspecting representative outputs from each cluster. The colour-coded clusters are then viewed in the input parameter space via linked SPloM — this is the parameter space partitioning, revealing which input regions produce each type of output behavior.
The user assesses the partitioning: whether clusters are meaningful and well-separated, whether the ROI needs narrowing (to focus on interesting regions), whether additional sampling is needed (to resolve boundaries), or whether new derived variables would improve cluster separation. If refinement is needed, they adjust the ROI, resample at higher density, compute additional features, recompute the embedding, and recluster. This cycle repeats — progressively building understanding of the parameter-output relationship — until the partitioning satisfactorily captures all qualitatively distinct behaviors.
From the final partitioning, the user extracts knowledge: parameter sensitivity (narrow clusters = high sensitivity), robust operating regions (large clusters tolerant to parameter changes), context-dependent optima (different "good" regions for different conditions), model comparison results (how two model versions differ in their partitioning), and identification of physically implausible or non-convergent regions to avoid.