Deep learning-aided simultaneous missing well log prediction in multiple stratigraphic units: a case study from the Bhogpara oil field, Upper Assam, Northeast India

Abstract

Accurate well log data is critical for subsurface characterisation and decision-making in the petroleum exploration. We explore and compare the effectiveness of three distinct deep leaning (DL) approaches—Long Short-Term Memory, Bidirectional Long Short-Term Memory, and Convolutional Long Short-Term Memory networks—in predicting missing well log data, a common challenge in the data acquired by Energy and Production (E&P) companies. Our analysis revealed the complex, nonlinear relationships present in geophysical logs through correlation matrix and determining the rank of predictor features through Minimum Redundancy Maximum Relevance (MRMR) analysis. To weigh these models, we used real-field wireline log datasets from the Bhogpara oil field of Upper Assam basin. The performance of each model is evaluated through root mean square error, correlation coefficients, mean absolute error and variance between actual and predicted values. The uncertainty of the models was facilitated by Monte Carlo simulation. Deep learning models accurately predicted neutron porosity logs from gamma-ray, resistivity, density, and photoelectric factor logs. The high correlation coefficients during the training (exceeding 0.90) and test (exceeding 0.97) phases illustrated the predictive precision of the DL models. Conv-LSTM consistently outperforms LSTM and Bi-LSTM, indicating the integration of convolutional layers in feature extraction offers a significant advantage in capturing intricate patterns in log data. The research showcases the effectiveness of deep learning architectures in predicting missing logs, a crucial aspect for E&P companies, as log data is vital for decision-making. The study presents a novel method for preserving data integrity and facilitating informed decision-making.