Geração de Fácies Geológicas a partir de um Modelo de Difusão Latente: Um Estudo Comparativo de Embeddings

Abstract

Geological facies are rock units with similar depositional characte ristics, essential for modeling subsurface heterogeneity in applications such as petroleum reservoir simulation and geological CO2 storage. In scenarios with limited real data, synthetic generation becomes a useful alternative. This study investigates the generation of 2D synthetic geological facies using a Latent Dif fusion Model (LDM) combined with different embedding strategies: Variational Autoencoder (VAE), VAE-GAN,andVector Quantized VAE(VQ-VAE).Eachem bedder wastrained under the same conditions and integrated into a common dif fusion pipeline. The models were evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Fr´echet Inception Distance (FID), and Learned Perceptual Image Patch Similarity (LPIPS). Among the strategies, the VAE achi eved the best quantitative performance: MAE = 0.0214, MSE = 0.0092, FID = 65.0449 and LPIPS=0.5643. Theotherapproachesshowedhighererrorvalues but exhibited characteristics suggesting potential for improved visual detail with extended training. The findings highlight a trade-off between model complexity and training efficiency in synthetic geological modeling.