home / skills / a5c-ai / babysitter / sensitivity-analysis-uq

sensitivity-analysis-uq skill

safe

/plugins/babysitter/skills/babysit/process/specializations/domains/science/mathematics/skills/sensitivity-analysis-uq

This skill enables global sensitivity analysis to quantify input importance and model behavior using Sobol, Morris, FAST, and DGSM methods.

npx playbooks add skill a5c-ai/babysitter --skill sensitivity-analysis-uq

Review the files below or copy the command above to add this skill to your agents.

Files (1)

SKILL.md

935 B

---
name: sensitivity-analysis-uq
description: Global sensitivity analysis methods
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit
  - Glob
  - Grep
metadata:
  specialization: mathematics
  domain: science
  category: uncertainty-quantification
  phase: 6
---

# Sensitivity Analysis (UQ)

## Purpose

Provides global sensitivity analysis methods for understanding model behavior and input importance.

## Capabilities

- Sobol indices computation
- Morris screening method
- FAST (Fourier amplitude sensitivity test)
- Correlation-based sensitivity
- Derivative-based sensitivity (DGSM)
- Variance-based decomposition

## Usage Guidelines

1. **Method Selection**: Choose based on computational budget
2. **Input Ranges**: Define appropriate input ranges
3. **Sample Size**: Ensure sufficient samples for convergence
4. **Interpretation**: Correctly interpret sensitivity indices

## Tools/Libraries

- SALib
- OpenTURNS
- UQLab

Overview

This skill provides a suite of global sensitivity analysis methods to quantify how model inputs drive output variability and to prioritize factors for uncertainty quantification. It wraps standard algorithms (Sobol, Morris, FAST, correlation- and derivative-based metrics) into a consistent API suitable for automated workflows. It’s designed to fit into deterministic, resumable orchestration pipelines and to scale with sampling budgets.

How this skill works

The skill generates input samples over specified ranges, runs the target model or simulator, and computes sensitivity measures from the resulting input–output dataset. It supports variance-based decomposition (Sobol), screening (Morris), frequency-domain testing (FAST), correlation metrics, and derivative-based global sensitivity (DGSM). Results are returned as indices and diagnostic statistics for ranking factors and assessing convergence.

When to use it

When you need to rank input importance across a high-dimensional model
During model validation to identify non-influential variables for simplification
To allocate computational budget by focusing on influential parameters
When preparing uncertainty quantification for decision-making or reporting
For screening prior to more expensive, detailed analyses

Best practices

Choose the method to match computational budget: Morris for cheap screening, Sobol for accurate variance decomposition
Define realistic input ranges and distributions before sampling
Ensure sample sizes are large enough for index convergence and run convergence diagnostics
Use correlated sampling or transforms if inputs are not independent
Compare multiple metrics (e.g., Sobol and DGSM) to validate conclusions

Example use cases

Ranking physical model parameters to reduce calibration dimensionality
Screening hundreds of inputs in a surrogate-model workflow
Assessing input contributions to risk metrics in engineering design
Comparing sensitivity under different scenario assumptions
Automated pipelines that trigger focused sampling based on preliminary screening

FAQ

Which method should I pick when I have limited compute?

Use Morris screening for a low-cost overview to identify non-influential inputs, then apply a variance-based method (Sobol) only to shortlisted variables.

Do inputs need to be independent?

Many methods assume independence; if inputs are dependent, use sampling transforms or methods that accept joint distributions and report results accordingly.