Sensitivity Analysis

Sensitivity analysis tells you which parameters the fit actually constrains and which it doesn’t. The standard tool is variance-based SOBOL indices over the parameter ranges. For the math (first-order vs total-order, Saltelli sampling), see the Sensitivity Analysis Guide.

Sensitivity analysis is currently a post-fit, in-process capability available through ionworkspipeline directly. It is not yet exposed as a separate ionworks-schema element or an ionworks-api endpoint — track pipeline.docs.ionworks.com for the standalone interface.

Interpreting the indices

After a fit, sensitivity analysis returns first-order (S1) and total-order (ST) indices plus 95% confidence intervals:

{
    "S1": [0.65, 0.25],
    "S1_conf": [0.05, 0.04],
    "ST": [0.72, 0.35],
    "ST_conf": [0.06, 0.05],
}

Quantity	Reading
`S1` near 1	The parameter alone accounts for most output variance
`ST − S1` large	The parameter’s effect is mostly through interactions
`ST` near 0	The parameter is unidentifiable inside its bounds — consider fixing it
`1 − sum(S1)`	Variance explained only by joint interactions

Workflow

Screen parameters

Start with n_samples=256 to identify the parameters with non-negligible ST (typically ST > 0.05).

Refine

For the parameters that matter, rerun with n_samples=1024 to tighten the confidence intervals.

Investigate interactions

Enable second-order indices when ST − S1 is large for multiple parameters — S2[i, j] reveals which pairs are coupled.

Pick n_samples as a power of two; the underlying SOBOL scheme uses

n \cdot (2p + 2)

model evaluations, where

p

is the number of fit parameters.

When too many evaluations fail

If a meaningful fraction of evaluations (>5%) fail because the model can’t solve the proposed parameter combinations:

Check that the parameter bounds are physically realistic
Verify the model is stable across the parameter space
Tighten bounds around the fitted values