Optimization of deepwater channel seismic reservoir characterization using seismic attributes and machine learning

Abstract

Accurate subsurface reservoir mapping is essential for resource exploration. In uncalibrated basins, seismic data, often limited by resolution, frequency, quality, etc., algorithms become the primary information source due to the unavailability of well logs and core data. Seismic attributes, while integral for understanding subsurface structures, visually limit interpreters to working with only three of them at once. Conversely, machine learning, though capable of handling numerous attributes, is often seen as inscrutable "black boxes," complicating the interpretation of their predictions and uncertainties. To address these challenges, a comprehensive approach was undertaken, involving a detailed 3D model from Chilean Patagonia's Tres Pasos Formation with synthetic seismic data. The synthetic data served as a benchmark for conducting sensitivity analysis on seismic attributes, offering insights for parameter and workflow optimization. The study also evaluated the uncertainty in unsupervised and supervised machine learning for deepwater facies prediction through qualitative and quantitative assessments. Study key findings include: 1) High-frequency data and smaller analysis windows provide clearer channel images, while low-frequency data and larger windows create composite appearances, particularly in small stratigraphic features. 2) GTM and SOM exhibited similar performance, with error rates around 2% for predominant facies but significantly higher for individual channel-related facies. This suggests that unbalanced data results in higher errors for minor facies and that a reduction in clusters or a simplified model may better represent reservoir versus non-reservoir facies. 3) Resolution and data distribution significantly impact predictability, leading to non-uniqueness in cluster generation, which applies to supervised models as well. Strengthening the argument that understanding the limitations of seismic data is crucial. 4) Uncertainty in seismic facies prediction is influenced by factors such as training attribute selection, original facies proportions (e.g., imbalanced data, variable errors, and data quality). While optimized random forests achieved an 80% accuracy rate, validation accuracy was lower, emphasizing the need to address uncertainties and their role in interpretation. Overall, the utilization of ground truth seismic data derived from outcrops offers valuable insights into the strengths and challenges of machine learning in subsurface applications, where accurate predictions are critical for decision-making and safety in the energy sector.