Reproducibility data and code for provenance-aware pre-build qualification of waste-derived AA6061 hybrid composite systems
Description
This dataset supports the study “Provenance-Aware Explainable Machine Learning for Resource-Efficient Pre-Build Qualification of Waste-Derived AA6061 Hybrid Composite Systems: A Transfer Framework for Metal Additive Manufacturing.” It contains source and processed data for six experimental response domains: compressive strength, tensile strength, flexural strength, wear rate, coefficient of friction, and fatigue cycles. The package includes provenance and formula audits, leakage-controlled machine-learning outputs, uncertainty diagnostics, explainability results, Pareto measurement-channel analyses, software-environment information, documentation, and integrity checks. The archive was designed to support transparent verification and computational reproducibility. Particular attention is given to identifying target proxies, formula-derived variables, redundant transformations, and other sources of data leakage before model development. The dataset is intended for pre-build materials qualification and methodological reproducibility, not for direct additive-manufacturing process optimization or inference of unmeasured AM process variables.
Files
Steps to reproduce
Use Python 3.13.5 with the software versions documented in the package, including NumPy 2.3.5, pandas 2.2.3, SciPy 1.17.0, scikit-learn 1.8.0, openpyxl 3.1.5, Matplotlib 3.10.8, and SHAP 0.50.0. Verify the integrity of the frozen source workbook using the supplied SHA-256 checksum. Read the workbook separately in formula-aware and cached-value modes. Do not recalculate spreadsheet formulas when generating the static processed CSV files, particularly the flexural-strength RAND() cell and the formula-derived fatigue variables. Run the provenance audit before model fitting. Identify constants, exact or near-exact redundant transformations, direct target proxies, and target-derived formula fields. Exclude inadmissible variables from the audited modelling feature sets. Reproduce the machine-learning benchmark using the fixed Ridge, PLS, SVR, Extra Trees, and mean-baseline configurations. Evaluate models using five repetitions of shuffled five-fold cross-validation with fixed random seeds, with preprocessing fitted only within the training folds. Reproduce the explainability, uncertainty, and measurement-burden analyses. Calculate held-out permutation importance and SHAP attribution, perform leave-one-out Gaussian-process uncertainty analysis, and evaluate all non-empty audited feature subsets for the Pareto analysis. Regenerate the reported tables, figures, out-of-fold predictions, uncertainty outputs, SHAP results, and Pareto results using the scripts supplied in the repository. Compare the regenerated outputs with the archived machine-readable results. The repository README.md provides the intended execution order and interpretation notes.