Synthetic tabular augmentation for peak particle velocity prediction in rock blasting

Synthetic augmentation can alter the bias and variance of a blasting-vibration predictor, but its benefit must be evaluated on independent real observations. This study examines a small peak particle velocity benchmark comparing Gaussian Copula, CTGAN, TVAE and CopulaGAN generators. The underlying database contained 88 records and was reduced to 80 modelling rows; 56 were used for training and 24 for real hold-out testing. The selected Gaussian Copula produced 537 retained synthetic records, giving 593 rows in the hybrid training table. It had the strongest reported marginal and pairwise resemblance, although the combined distance and rank summaries are descriptive rather than inferential tests. On the fixed real test set, hybrid training reduced Random Forest RMSE from 3.44 to 2.79 mm/s and XGBoost RMSE from 3.38 to 2.93 mm/s. These changes correspond to reductions of 18.9% and 13.3%. Real-only Ridge remained the strongest reported test predictor, with RMSE = 2.05 mm/s and R² = 0.89. The evidence therefore supports model-specific gains within one split, rather than a universal advantage for augmentation. Interpretation is limited by pre-split cleaning, incomplete documentation of generator refitting within cross-validation and the absence of repeated independent real tests. The discussion connects these limits to charge-distance physics, log-scale retransformation, synthetic-data weighting and upper-tail vibration risk, and defines a validation design that can distinguish useful regularisation from optimistic evaluation.