Leveraging Neural Radiance Fields for High-Fidelity 3D Digital Outcrop Reconstruction and Characterization

Abstract This study continues SPE-216868-MS and adds a novel application of Neural Radiance Fields (NeRF) for digital outcrop modeling and rock characterization. NeRF as an open-source AI/ML framework, enables the synthesis of high-fidelity 3D reconstructions from 2D images. We leverage NeRF to create the high-fidelity, real world 3D scenes of the rocks and outcrops. We use a combination of field observations, photographs, and LiDAR scans to create a large dataset of outcrop samples from various locations. Our methodology consists of three main steps: data preprocessing, training an optimized NeRF model, and analysis of the resulting outputs. In the first step, we preprocess and align the collected imaging data to ensure consistent lighting conditions and remove any noise and artifacts. Next, we train the NeRF model using the preprocessed data, which enables us to reconstruct the 3D geometry and appearance of the outcrops. Finally, we analyze the trained models to extract rock features, such as texture, porosity, permeability, and mineralogy. Multi-view geometry and incorporated physical constraints reconstruct the 3D geometry and appearance of the rock sample. This use case and the workflow above demonstrated how the subsurface data can be made accessible to the broad audience, independent of its technical proficiency or familiarity with the specific region of the data itself. The whole workflow is designed with almost no coding, chaining together open source tools with easy-to-use no-code interfaces on the cloud. Once the initial solution is setup, it is also easily extensible to additional scenarios and data types. The solution saves engineer’s time and resources and leads to more accurate geological interpretations and increased exploration success. By integrating multi-view geometry and incorporating physical constraints, our method faithfully reconstructs the 3D geometry and appearance of the rock sample, capturing intricate features at sub-millimeter scales. This rock models can be taken further to the numerical simulations or AI/ML models to calculate various reservoir properties such as fluid saturation, pore pressure, and more.