Prediction profiler
Drag a factor, watch the prediction move everywhere.
A prediction profiler turns a fitted model into something you can steer. Drag any factor and watch the predicted response, and its trace in every other panel, update at once, so you can read interactions, find settings, and let a desirability function optimize across several responses together.
What it measures
Once a model is fit, whether from a factorial, a response surface, or a Gaussian process, the profiler is how you interrogate it:
- Linked partial-dependence panels: one panel per factor, each showing how the predicted response changes as that factor moves with the others held at their current cursor settings. The cursors are shared, so the panels are a single coordinated view of one point in factor space.
- Live interaction reading: move one factor's cursor and the slope of another factor's panel changes, that change is an interaction, made visible and tangible instead of inferred from a coefficient table.
- Desirability optimization: when there are several responses to satisfy at once, maximize capacity, minimize resistance, keep thickness on target, each gets a desirability function, and the profiler finds the factor settings that maximize the combined desirability, the multi-objective optimum.
How to read the output
Use the profiler to turn a fitted model into decisions. Sweep a cursor and read the slope: steep means the response is sensitive to that factor here, flat means it is not, and a slope that changes as you move a different factor is the interaction you need to respect when you set the process. For multiple responses, watch the trade-offs directly, pushing one factor to raise capacity may lift resistance past its target, and the desirability optimum is the balance point the profiler settles on. The caution is the model's reach: a profiler will happily predict anywhere, including outside the region your data supports, so trust the readings inside the design space and treat extrapolation, especially with a response-surface model, as a hypothesis to test, not a result.
A real use case
After a response-surface study on a cathode coating, a team has a model relating solids content, coating gap, and line speed to three responses: areal loading, adhesion, and defect rate. The numbers are in a coefficient table nobody can feel. Dropped into a prediction profiler, the model becomes navigable: dragging line speed up shows defect rate climbing steeply past a threshold while loading barely moves, and the adhesion panel's slope flips sign depending on solids content, an interaction the table buried. With desirability set to hit the loading target, keep adhesion above spec, and minimize defects, the profiler lands on a setting the team would not have guessed, moderate speed with higher solids, and they validate it on the line instead of running another design.
Common mistakes
- Extrapolating outside the design region. The profiler predicts anywhere, but the model only has support where you collected data.
- Reading one panel in isolation. The panels are linked; a factor's effect can change completely when another factor's cursor moves, that is the interaction.
- Optimizing one response and ignoring the others, when the real problem is a multi-response trade-off a desirability function is built to balance.
- Trusting a confident-looking prediction from a model that fit poorly. The profiler only reflects the model; check the fit quality before believing the surface.
- Setting desirability weights without thought, so the optimum reflects an unintended priority rather than the real engineering goal.
Make the fitted model navigable, then optimize it
Niobia presents the fitted model, from a factorial, central composite, Box-Behnken, or surrogate fit, as a prediction profiler with linked partial-dependence panels, so changing any factor updates the predicted response and its trace in every panel at once and the interactions become visible. For multi-response problems it applies desirability functions across the responses and finds the settings that maximize the combined desirability. It is the interactive companion to Bayesian optimization: the profiler is for understanding and steering the model by hand, Bayesian optimization is for letting the agent choose the next run, and both read from the same fitted surface.
Frequently asked
What does a prediction profiler actually show?
How the fitted model's predicted response changes as you move each factor, across linked panels that share cursors. Moving one factor updates the prediction in every panel, which makes the model's behavior, including interactions, explorable by hand instead of read from a coefficient table.
How does it handle multiple responses at once?
Through desirability functions: each response (maximize, minimize, or hit a target) is scored, and the profiler finds the factor settings that maximize the combined desirability. That is the multi-objective optimum, the best balance across responses that often conflict.
Can I trust predictions outside the range I tested?
Treat them as hypotheses, not results. The profiler will predict anywhere, but the underlying model only has data support inside the design region. Extrapolation, especially with a response-surface model, should be validated with a confirming run before you rely on it.