Abstract Subsurface modeling is essential for characterizing resources such as hydrocarbon reservoirs, aquifers, mining ore bodies, and carbon sequestration sites, which is crucial for resource management amid uncertainty. Traditional geostatistical methods, while effective, face challenges in capturing the complexity of subsurface features, particularly in deepwater depositional systems where geological heterogeneity is significant. Recent advancements in machine learning (ML), including generative adversarial networks (GANs) and conditional GANs techniques, have shown promise in improving subsurface models by incorporating geological information and spatial patterns. However, the performance of ML models depends on high-quality training datasets, which are currently limited, especially for deepwater settings. To address this, we propose an open-source Python package, GeoRulesLobePy, for generating rule-based deepwater training images that integrate geological observations with Markov and morpho-dynamic rules. This package is designed to be user-friendly, computationally efficient, and capable of producing realistic geological models. GeoRulesLobePy simulates deepwater depositional processes by sequentially placing lobe elements within a 3D grid, following rules for geometry, stacking patterns, and facies trends. The methodology ensures that the generated models capture the hierarchical and heterogeneous nature of deepwater lobe complexes, making them suitable for training ML algorithms. This package was tested using scenarios from the Golo system in Corsica, France, and the Tanqua Karoo basin in South Africa, demonstrating its versatility and capability to produce realistic subsurface models. By providing an open-source tool for generating complex, rule-based training datasets, GeoRulesLobePy offers a tool for geoscientist to visualize their 2D observations in 3D, and creates a vast training data set for ML purposes..
Chacon-Buitrago et al. (Thu,) studied this question.