The Hybrid Visual Steering workflow supports designers of complex engineering systems (here, a common-rail Diesel injection system) in exploring large simulation ensembles to find optimal control-parameter settings. Its central idea is to tightly couple three components that are normally used in isolation — simulation, interactive visual analysis, and automatic optimization — within a single integrated design environment, so that automatic optimization guides the analyst while the analyst steers where new simulation runs are computed.
The process begins by creating an initial design of experiment (DOE), sampling a limited, discrete set of control data points from the high-dimensional parameter space. The simulation is then run for each sampled point, producing complex time-series outputs (families of curves over the crankshaft cycle). Because regression and optimization require scalars, a feature-extraction step replaces each curve with one or more scalar features. Two coordinated overview visualisations are produced: the Parameters Exploration View (stacked per-parameter histograms with constraint bars) and a set of result views (scatterplots, parallel coordinates, and family-of-curves views), all linked through brushing.
The analyst then enters an iterative hybrid steering loop. By brushing across the coordinated views they study parameter sensitivity, reconstruct parameters for desired outputs and curve shapes, and compare scenarios — identifying a promising region of interest. They specify optimization constraints and goals via the constraint bars, after which a regression surrogate model is built from all runs (or a brushed subset). The Regression Exploration View projects the solution space onto two output values and overlays simulation and regression points with deviation lines, making the model's accuracy explicit. Automatic optimization over the regression model, subject to the constraints, determines a candidate optimal control point.
The analyst assesses whether this optimum is trustworthy — checking regression accuracy in the region, whether better neighbouring points exist, and whether the resulting curves have acceptable shapes. If the result is unsatisfactory, a new DOE with increased resolution is created around the optimum, the new runs are simulated and their features extracted, the runs are integrated into the growing ensemble, and the views are re-rendered (now also showing the computed optimum). The loop repeats on the augmented ensemble until a satisfactory optimum is reached.
Finally, in a decision step, the analyst selects the optimal control parameters, confirms acceptable curve shapes, and records design insights and any need for additional simulation runs.