Common-basis analysis of three-dimensional ECT sensor and measurement designs

Version 1.1.0 contains the common-basis analysis and additional numerical studies used in the revised manuscript:

Common-basis computational analysis reveals task-dependent trade-offs in three-dimensional electrical capacitance tomography sensor and measurement design

Author: Robert Banasiak, Lodz University of Technology.

Included analyses

• Primary volume-integrated analysis on the fixed 2851-bin domain, spectral and spatial analysis, 128-channel selection, 3000 matched reconstructions and 630 paired statistical comparisons.
• Mapping/grid sensitivity with Q = 1, 64 and 512 on two cylindrical grids.
• Selection baselines and sensitivity to regularisation, subset size and noise.
• Runtime and peak-process-memory benchmark of the primary pipeline.
• A-C cross-mesh reconstruction: 36 cases and 5760 reconstructions.
• Original reference results, manifests, source provenance and file checksums.

These are computational comparisons of precomputed FEM models. The primary reconstruction benchmark uses matched linear forward and inverse operators. Subdivision quadrature approximates integration of native piecewise-constant sensitivity; it is not an exact curved-bin intersection calculation. The A-C experiment is a linear cross-mesh test, not experimental or independent nonlinear validation.

Native inputs

The fifteen native matrices are archived separately at:

https://doi.org/10.5281/zenodo.21409630

Place meshA.mat through meshO.mat in matrices/, or pass their directory with --input-dir. These large inputs are not duplicated in this distribution. The supplied geometry uses millimetres and the sensitivities are densities.

Environment

Use Python 3.12.4 with the libraries pinned in requirements.txt. For example on Windows, select a Python 3.12.4 installation:

py -3.12 -m venv .venv
.venv\Scripts\activate
python -m pip install -r requirements-dev.txt

On Linux/macOS use python3.12 -m venv .venv and source .venv/bin/activate before installing the requirements. Run the commands below from this release's root directory.

Checks without native matrices

python -m pytest -q
python run_q512_revision.py --self-test
python run_q512_selection_revision.py --self-test
python run_q512_benchmark_revision.py --self-test
python run_q512_model_mismatch.py --self-test

These are synthetic numerical and integration checks, not a replacement for rerunning the full native-input experiments.

Primary analysis

python run_q512_revision.py --input-dir matrices

Use run_q512_revision.py for the revised primary analysis. main.py is retained as the original centroid-mapping configuration and backend; running it alone uses the older mapping. The primary entry point creates a fresh q512_revision_results/run_* directory, uses the frozen original 2851-bin mask, recalibrates the shared alpha once on the mapped operators, and verifies full-precision CSV exports. Native inputs and previous runs are not overwritten.

Additional studies

After the primary run completes:

python run_q512_selection_revision.py --selection-study baselines --input-dir matrices
python run_q512_selection_revision.py --selection-study stability --input-dir matrices
python run_q512_benchmark_revision.py --input-dir matrices
python run_q512_model_mismatch.py --input-dir matrices

These entry points find the latest complete primary run under q512_revision_results/. Use --primary-run PATH to select a particular run. Keep the source files unchanged between the primary run and its dependent studies: both source and native-input hashes are checked.

The publication mapping-sensitivity settings override the helper's defaults:

python analyze_mapping_sensitivity.py --input-dir matrices --mapping-grids 10x16x20 20x32x40 --subdivision-levels 0 2 3

Levels 0, 2 and 3 correspond to Q = 1, 64 and 512. This study holds the centroid-reference regularisation density fixed; it is separate from the recalibrated primary analysis. Figure 1 can then be generated from compatible completed runs with python plot_fig1_q512.py.

Reference results and provenance

reference_results/ contains the author's six completed numerical runs grouped by study, with original tables, figures and manifests. reference_results/INDEX.json records the original archive names and SHA-256 hashes. Historical local paths and recorded environments in those manifests have been preserved.

The numerical source files match the hashes recorded in the reference runs. Only the version string in ect_analysis/__init__.py has changed to 1.1.0; this metadata-only difference is recorded in provenance/source_audit.json. Generate a fresh primary run from this release before running its dependent studies so their source hashes include the current release metadata. Reference outputs are kept separate from automatically discovered new runs.

Read exported float64 tables using pandas.read_csv(path, float_precision="round_trip"). SHA256SUMS.txt covers the distributed files, and local checks are recorded in provenance/release_validation.json. Further settings and limitations are in docs/REPRODUCING.md. A guide to the manuscript results and their source tables is in docs/MANUSCRIPT_ALIGNMENT.md. The journal-ready paired-statistics CSV is included as supplementary/Supplementary_Data_S1_pairwise_statistics.csv.

Citation metadata: CITATION.cff. Software publication instructions: docs/RELEASING.md. The software and input dataset have separate citations. License: BSD-3-Clause; see LICENSE.