digiqual

Statistical Toolkit for Reliability Assessment in NDT

digiqual is a Python library designed for Non-Destructive Evaluation (NDE) engineers. It implements the Generalised $\hat{a}$-versus-a Method, allowing users to perform reliability assessments without the rigid assumptions of linearity or constant variance found in standard methods.

Documentation: Read the full documentation here

Installation

Just want the app? Windows and macOS installers are attached to the latest GitHub Release; no Python needed. See Launch the App for install notes.

You can install digiqual directly from PyPI.

Option 1: Install via uv (Recommended)

If you are managing a project with uv, add digiqual as a dependency:

# To install the latest stable release (v0.28.0):

uv add digiqual

# To install the latest development version (main branch from github):

uv add "digiqual @ git+https://github.com/JGIBristol/digiqual.git"

If you just want to install it into a virtual environment without modifying a project file (e.g., for a quick script), use pip interface:

uv pip install digiqual

Option 2: Install via standard pip

To install the latest stable release (v0.28.0):

pip install digiqual

To install the latest development version from github:

pip install "git+https://github.com/JGIBristol/digiqual.git"

Features

1. Experimental Design

Before running expensive Finite Element (FE) simulations, digiqual helps you design your experiment efficiently.

• Latin Hypercube Sampling (LHS): Generate space-filling designs over your parameter ranges, covering both the parameters of interest (e.g. defect size) and nuisance parameters (e.g. roughness, orientation).
• Run in any order, stop early: Points are reordered with a greedy max-min algorithm, so any first k runs are already well spread.

2. Data Validation & Diagnostics

Ensure your simulation outputs are statistically valid before processing.

• Sanity Checks: Detects overlap between variables, type errors, and insufficient sample sizes.
• Sufficiency Diagnostics: Five tests (input coverage gaps, model fit, two bootstrap-stability checks and collinearity/VIF) flag problems before you trust the results.

3. Adaptive Refinement (Active Learning)

digiqual closes the loop between analysis and design.

• Smart Refinement: Use refine() to identify specific weaknesses in your data. It uses bootstrap committees to find regions of high uncertainty and suggests new points exactly where the model is "confused".

• Automated Workflows: Use the optimise() method to run a fully automated "Active Learning" loop. It generates an initial design, executes your external solver, checks diagnostics, and iteratively refines the model until statistical requirements are met.

4. Generalised Reliability Analysis

The package includes a full statistical engine for calculating Probability of Detection (PoD) curves.

• Relaxed Assumptions: Moves beyond the rigid constraints of the classical $\hat{a}$-versus-$a$ method by handling non-linear signal responses and heteroscedastic noise.
• Multi-Dimensional Active Marginalisation: Resolves multidimensional physics by integrating out stochastic nuisance parameters (like roughness) via Monte Carlo methods, outputting high-fidelity 2D PoD surface heatmaps alongside standard 1D curves.
• Model Selection: Compares polynomials (degree 1–10) and a Kriging (Gaussian Process) model by 10-fold cross-validation, and picks the simplest model within one standard error of the best.
• Kriging Metamodelling: Anisotropic Matérn/Gaussian kernels with a learned noise term, kernel choice by leave-one-out error and LOO residual diagnostics, following Malkiel et al. (2026).
• Varying Scatter and Non-Gaussian Errors: Models how the scatter changes across the inputs with a kernel smoother, and picks the error distribution (e.g. Normal, Gumbel, Logistic) by AIC.
• Uncertainty Quantification: Bootstrap resampling gives confidence bounds at several levels at once, and a full $a_{X/Y}$ reliability matrix (e.g. $a_{90/95}$).
• Sensitivity Analysis: Total-order Sobol indices show which inputs drive the signal.
• Classical Comparison: linear_pod() runs the classical $\hat{a}$-versus-$a$ analysis on the same data for side-by-side comparison.

5. Speed

• Layered caching: Models are fitted once, so changing a threshold or slice afterwards is near-instant.
• C++ acceleration: The kernel smoother and Monte Carlo integration run in a multi-threaded C++ extension, with an automatic NumPy fallback.

6. Desktop and Browser App

The full workflow, from experimental design to UQ, is also available as a point-and-click app: a Windows/macOS desktop application, or uvx digiqual from a terminal.

Development

If you want to contribute to digiqual or run the test suite locally, follow these steps.

1. Clone and Install

This project uses uv for dependency management.

git clone https://github.com/JGIBristol/digiqual.git
cd digiqual

2. Run Tests and Lint

The package includes a full test suite using pytest. Development tools are uv dependency groups, installed automatically by uv run.

uv run pytest            # or: just test  (the memory stress test is opt-in: just test_stress)
just lint                # ruff

3. Build Documentation

To preview the documentation site locally:

just preview

4. Work on the App

The GUI is the digiqual.gui subpackage (src/digiqual/gui/); app/ only holds the Briefcase configuration for the Windows/macOS installers. Run it with just app (desktop window) or just app_dev (browser, live reload). See Desktop App Architecture for how it fits together and how the installers are built.

References

Malkiel, N., Croxford, A. J., & Wilcox, P. D. (2025). A generalized method for the reliability assessment of safety–critical inspection. Proceedings of the Royal Society A, 481: 20240654. https://doi.org/10.1098/rspa.2024.0654

Malkiel, N., Croxford, A. J., & Wilcox, P. D. (2026). A comprehensive investigation of flexible and multi-dimensional simulation-based PoD analysis. NDT & E International, 159: 103596. https://doi.org/10.1016/j.ndteint.2025.103596