Deep learning-based prediction of effective offshore reservoirs using multimodal and multi-scale data in highly heterogeneous and fluvial environments

This paper introduces a sophisticated ensemble model for reservoir characterization that integrates multimodal datasets to reduce uncertainties and improve prediction accuracy. The model combines results from three separate models: (1) a deep learning model using 1D well logs, (2) a deep learning model using 2D/3D seismic data, and (3) a constraint model that refines predictions using additional data such as geo conceptual model. The design follows a dynamic knowledge build-up approach, allowing predictions to be updated dynamically when new data of the same or different types are introduced. For example, the model can predict using only 1D well logs and update its predictions easily when new well logs or seismic data become available. The final reservoir prediction is provided as an effective reservoir in 3D grid, with the model capable of generating hundred million-level predictions in just a few minutes. This efficiency extends to result updating when new data is incorporated. The model addresses key challenges in reservoir characterization, including data scale differences, data scarcity, and multimodal data integration, offering a robust and efficient solution for dynamic reservoir prediction.