Machine Learning Based Automatic Marker Clustering

Abstract

Abstract Accurate information pertaining to the markers or well tops is critical to define the key formation boundaries in the regional stratigraphic frameworks. The information with respect to the markers in larger basins is spread across multiple sources as the various tools utilized in the Oil and Gas industry, public data repositories and the well logs provided by data vendors. Collaboration across such multiple sources of information is essential if we need to develop subsurface models for the larger basins. However, such collaboration is non-trivial and is accompanied by challenges. The well tops information is interpreted uniquely by interpreters across different organizations, probably using different stratigraphic columns. The multitude of well top information sources, interpreters and stratigraphic columns introduce inconsistencies in the well top information. We propose deep learning based unique interpretation of the various marker log patterns which can consistently map them across the wells in a region. The transformer model is trained using a self-supervised learning approach to predict the randomly masked well log values in the input. This allows the transformers to learn the feature hierarchy in the marker log patterns which is manifested in the latent space representations. The proposed approach utilizes transformer latent space to represent markers in a lower dimensional feature space which are further projected using Uniform Manifold Approximation and Projection (UMAP) to two dimensional embeddings. We apply Gaussian Mixture Models (GMM) based probabilistic clustering on the embeddings to automatically separate the marker representations. We evaluate the transformer latent space based automatic marker separation models using Bakken open-source dataset and choose six markers for the evaluation. We find that the proposed approach can separate the markers with very high accuracy, giving us excellent precision and recall for all analyzed markers. This approach enables automatic separation of the marker patterns and maps them consistently across the wells in a basin.