An Innovative Method for Evaluating the Fracturability of Continental Shale Oil Reservoirs Based on Machine Learning and Logging Data

Summary Fracturability evaluation is an important task before hydraulic fracturing, and machine learning (ML) methods have been applied to petroleum-related studies but have not been applied to fracturability evaluation. In this study, we present a novel fracturability evaluation workflow based on ML that focuses on the Chang 7 continental shale oil reservoirs in the Ordos Basin, aiming to generate a comprehensive fracturability index (KC) to guide the selection of sections and clusters in hydraulic fracturing. Owing to the lithological differences between continental shale oil reservoirs and shale reservoirs, brittleness-based fracturability evaluation methods are poorly adapted to the research area. Hence, we integrate geological sweet spots, brittleness, and difficulty in forming a complex fracture network in a reservoir. The typical factors influencing fracturability include oil saturation (SOG), porosity (POR), permeability (PERM), Young’s modulus (YM), Poisson’s ratio (PR), Mode-I fracture toughness (KIC), Mode-II fracture toughness (KIIC), and the horizontal stress difference coefficient (HSDC). Furthermore, the powerful nonlinear dimension reduction capability of kernel principal component analysis (KPCA) is used to integrate the main characteristics of each effect. To verify the adaptability of the KPCA-based method, the KC is compared with logging interpretations and microseismic events. Considering the substantial spatial correlation of logging data, a hybrid neural network [convolutional neural network (CNN)-multihead attention (MHA)-bidirectional long short-term memory neural network (BiLSTM)] is presented to simplify the complex intermediate computation procedure and directly use the logging data to predict the KC. The CNN excels at extracting local features, the MHA enables the model to focus more on task-relevant data, and the BiLSTM captures bidirectional dependencies with logging data. The experimental results show that the CNN-MHA-BiLSTM outperforms other neural networks on the testing set and can better handle data with hidden patterns.