Strata-constrained GWLSTM network for logging lithology prediction

Abstract

Precisely identifying rock lithology from logging curves is critically important for reservoir characterization and exploration risk assessment. Although traditional knowledge-based lithology interpretation by the well logging interpreter has achieved success, the interpreter-dominated lithology prediction process, in turn, could lead to a biased prediction or erroneous decision-making. Deep neural network shows the most advanced performance in various domains such as medical science, computer vision, or even geosciences. Therefore, a potential strata-constrained long short-term memory (LSTM) strategy is proposed. By combining Gaussian windows to characterize the weighted stratigraphic sequence information on the target formation, rock lithology can be intelligently identified from the input logging curves. This weighted stratigraphic sequence constrains-based LSTM workflow can predict the rock lithology precisely, even for the thinner layers. F1 score and confusion matrix demonstrate that considering the rock strata sequence features, the predicted lithology by the Gaussian window weighted constrain LSTM (GWLSTM) model and rectangular constrain LSTM (RCLSTM) model have superior performance than those of conventional LSTM model.