The application of the CNN-BiLSTM-TPA model based on deep learning in porosity prediction

Abstract Porosity is a fundamental physical property in reservoir evaluation, and its prediction accuracy directly affects the scientific validity and reliability of reservoir characterization. However, traditional empirical methods often struggle to accurately capture the complex and nonlinear relationships among parameters, leading to significant prediction errors. In recent years, deep learning techniques have gained widespread application in geotechnical engineering, oil and gas exploration, and related fields due to their strong capabilities in modeling nonlinear dependencies. To improve the accuracy and generalization capability of porosity prediction, this study proposes a hybrid deep learning framework named convolutional neural network-bidirectional long short-term memory-temporal pattern attention (CNN-BiLSTM-TPA), which integrates multiple neural network architectures with a physical consistency constraint. The model combines the shallow spatial feature extraction capabilities of CNNs, the sequential dependency modeling strength of BiLSTM networks, and a multi-scale attention correction mechanism via the TPA module. This integrated design effectively addresses key challenges in geological modeling, such as nonlinearity, time lag, and localized periodicity. Empirical evaluations based on real logging data from the Zhujiangkou Basin in the South China Sea, as well as cross-regional validation using data from the Xinjiang Oilfield, demonstrate the proposed model’s superior performance in terms of prediction accuracy and robustness. These results highlight its strong engineering adaptability and broad application potential.