Deep learning based automatic marker separation
Atul Laxman Katole, Aria Abubakar, E. V. Hoekstra, Srikanth Ryali, Zhao Tao · 2023
Transitions in the geological layers are referred as markers and they define the key formation boundaries in the regional stratigraphic frameworks. The information pertaining to the markers is gathered from multiple sources as numerous tools prevalent in the Oil and Gas industry, multiple public data bases and vendor supplied datasets. The interpretation of these markers collected from the multiple sources lacks consistency due to the involvement of multiple personnels and different standards. The authors propose a deep learning-based approach to consistently map the markers across the wells in a basin. The proposed automatic marker separation approach is based on dimensionality reduction of the marker waveforms using latent space representation in the transformer models. The latent representation is further transformed to two-dimensional embeddings using Uniform Manifold Approximation and Projection (UMAP). The resultant two dimensional embeddings can be readily segregated using clustering approaches as K-Means and Gaussian Mixture Models (GMM). The probabilistic clustering approach as GMM provides the probability of marker assignment to a specific cluster thereby providing us the confidence in each assignment. The proposed automated clustering approach is evaluated using the six markers available in a well dataset from the Williston basin in the North America. We find that the proposed marker separation approach provides superior accuracy, giving us high precision and recall for all the analyzed markers. The proposed approach paves the way for automated marker separation which leads to consistent mapping of markers across hundreds of wells in a basin.