Input Normalization for WGAN-GP Synthesis of Radiometric Activity Data from Building Materials

Resumo This study evaluates the extent to which input normalization influences the performance of a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) trained on gamma-spectrometric measurements of 226 Ra, 232 Th, and 40 K activity concentrations, together with six UNSCEAR-derived radiological hazard indices, for commercial building materials from Rio de Janeiro, Brazil. Six candidate preprocessing pipelines were initially defined: linear MinMax scaling, logarithmic transformation followed by MinMax scaling, empirical quantile mapping to approximately Gaussian marginals followed by MinMax scaling, median-IQR robust scaling followed by MinMax scaling, a composite logarithmic-standarization-MinMax pipeline, and a skewness-guided adaptive procedure. Prior to any training, algebraic analysis established that two of these pipelines reduce to simpler ones under the adopted implementation, leaving four geometrically distinct preprocessing strategies. Forty training runs (four conditions, ten paired random seeds each) supported a randomized complete block design and non-parametric inference across six evaluation axes: adversarial stability, marginal and multivariate fidelity, manifold quality, structural preservation, physical consistency, and memorization risk. The results suggest that normalization may affect multivariate similarity more consistently than marginal reconstruction. Nonlinear transformations were generally associated with lower Maximum Mean Discrepancy (MMD) relative to linear scaling, whereas no pairwise difference in marginal Earth Mover's Distance (EMD) survived multiplicity correction. Quantile normalization was associated with a lower adversarial-collapse rate but also with reduced manifold recall and coverage under the fixed training budget, a pattern compatible with a stability-coverage trade-off. Uniform logarithmic transformation substantially degraded the 40 K marginal, a variable with an approximately symmetric distribution, supporting the use of variable-specific preprocessing when radiometric features exhibit heterogeneous distributional profiles. Violations of the deterministic relationships defining derived radiological indices persisted across all conditions and all six indices, suggesting that normalization alone may be insufficient to ensure physical consistency, and that constraint-aware generative architectures may be required when radiological compliance is a primary objective. No evidence of training-data memorization was identified under the adopted diagnostic framework.