Synthetic Well Log Generation with Preserved Multivariate Correlations and Vertical Facies Stacking Patterns
We present a novel procedure for generating synthetic well logs that simultaneously preserves multivariate correlations among petrophysical properties (Density, P-Sonic, S-Sonic) and vertical stacking patterns of electrofacies. The methodology integrates Markov chain models, autoencoder-based dimensionality reduction, and Markov chain Monte Carlo (MCMC) sampling in latent space. Application to a real turbidite reservoir dataset demonstrates that the framework successfully sustains fundamental rock physics relationships and generates geologically realistic vertical heterogeneity consistent with actual well log measurements. This technique addresses critical data scarcity in machine learning applications for seismic interpretation while enabling credible synthetic seismogram generation for scenario testing and uncertainty quantification in petroleum exploration and field development.