ARTIFICIAL INTELLIGENCE FOR LITHOFACIES IDENTIFICATION USING 1-D CONVOLUTIONAL NEURAL NETWORK

Abstract

Accurate lithofacies identification is an essential step in reservoir characterization. Traditional workflows depend on manual interpretation of core samples and well logs, which are time-consuming, expensive, and often influenced by interpreter bias. This study introduces a data-driven framework that applies one-dimensional convolutional neural networks (1-D CNNs) to automate lithofacies classification directly from wireline log measurements. The workflow begins with preprocessing of well logs, where redundant columns are removed, missing values are handled with sentinel flags, and the data are standardized to zero mean and unit variance. Feature importance analysis using an XGBoost classifier highlights gamma-ray, density, resistivity, and porosity-related logs as the most influential inputs for classification. Based on these insights, a compact 1-D CNN architecture is developed with three convolutional blocks, each including batch normalization, pooling, and dropout, followed by global average pooling and a softmax output layer. Model training is conducted with stratified three-fold cross-validation, early stopping, and learning-rate scheduling. The final model achieves approximately 90% accuracy and F1 scores close to 0.90 across more than 1.2 million depth samples. Learning curves show that the model converges in fewer than 20 epochs, while predicted facies logs demonstrate strong agreement with gamma-ray trends and known lithologic boundaries. This approach minimizes human bias, reduces interpretation time from weeks to hours, and generates reproducible facies logs that can be directly applied in reservoir modeling.