Abstract
High-quality synthetic datasets are critical for environmental monitoring due to the high cost of primary geochemical data collection. However, selection criteria for generative frameworks under small-sample constraints remain poorly defined. This study evaluates two architectures—the deep-learning-based Tabular Variational Autoencoder (TVAE) and the parametric Gaussian copula—using an empirical geochemical dataset (N = 297) of 25 log-transformed elements. Synthesis quality was benchmarked via marginal distribution fidelity (KSComplement), correlation preservation (CorrelationSimilarity), privacy (LogisticDetection), and a composite QualityScore, supplemented by divergence metrics and a rare-earth element (REE+Th+U) subgroup analysis. In this dataset, the Gaussian copula demonstrated superior global correlation preservation (CorrelationSimilarity: 0.9633 vs. 0.9482), particularly for weak-to-moderate dependencies. Conversely, TVAE better replicated marginal distributions (KSComplement: 0.8742 vs. 0.8596), maintained localized correlations (MAD: 0.0728 vs. 0.1196), and showed enhanced privacy (LogisticDetection: 0.5757 vs. 0.3063). These complementary profiles suggest that, for this case study, the Gaussian copula may be preferable for dependency modeling, while TVAE appears better suited for secure open-data dissemination. Further validation on additional datasets is needed to assess the generalizability of these findings.