This talk introduces a digital capability developed to accelerate creation of sub-surface velocity scenario generation and assessment, a time and effort-intensive albeit crucial part of the velocity model building workflow. Velocity scenario generation is a complex, data-intense and interpretive process and hence deep-learning (DL) is well-suited solution to accelerate it. However, due to the variability & complexity of scenario generation, the quality and quantity of data required to train DL models turned out to be extremely difficult to gather. We turned our focus to leveraging traditional computer vision techniques and developed a single set of building blocks to easily customize velocity scenario generation workflows for specialized projects and specific needs, enabling creating multiple velocity scenarios for various surveys in an interactive and user-guided workflow. Furthermore, we are not only able to address data scarcity as this user-guided workflow also serves as a high-quality training data generator for DL models that will gradually replace the computer-vision techniques enabling embedding domain knowledge into intelligent scenario generation capability.
Authors: Apurva Gala (Shell), Pandu Devarakota (Shell), Engin Alkan (Shell), John Kimbro (Shell) and Gislain Madiba (Shell)
