Deep Learning-Based Automated Workflow for Stratigraphic Boundary Detection and Well Correlation

Stratigraphic boundary detection and well-to-well correlation are critical for subsurface characterization and reservoir analysis. Traditional approaches depend heavily on manual interpretation, making them time-consuming, inconsistent, and error-prone—especially in the context of big data from large well datasets. This study introduces an automated deep learning-based workflow that combines an attention-driven model with the Hungarian algorithm for stratigraphic boundary detection, followed by a feature vector similarity-based correlation engine. The workflow is trained and evaluated on a large-scale dataset of approximately 2,000 wells from the Athabasca Oil Sands Area. Preprocessing included de-spiking, scaling, and segmenting logs into fixed-size patches to support batch processing and model efficiency. The model accurately detects boundary depths, achieving 96% accuracy, with precision, recall, and F1-scores of 0.85, 0.92, and 0.87, respectively. Detected boundaries are then used to segment logs into formation-like intervals, from which latent feature vectors are extracted. Formation correlation is performed using cosine similarity between vectors of reference and target wells. This automated and scalable pipeline significantly reduces manual effort and improves consistency in geological interpretation, making it well-suited for data-driven reservoir studies and other geoscientific applications requiring high-throughput analysis.