Facies classification using the convolutional neural network (CNN) algorithm in an offshore oilfield, SW of Iran

Accurate lithofacies classification is essential for effective reservoir characterisation and hydrocarbon development. This study presents a lightweight one-dimensional convolutional neural network (1D-CNN) for automated facies classification using well-log data from two offshore wells in south-western Iran. Four petrophysical logs, gamma ray, resistivity, sonic transit time, and bulk density served as input features to distinguish calcite, dolomite, and anhydrite facies. Five optimisation algorithms (Adagrad, Adadelta, Adam, Adamax, and stochastic gradient descent) were evaluated based on classification accuracy, convergence behaviour, and computational efficiency. The CNN architecture incorporates batch normalisation, dropout regularisation, and fine-tuned hyperparameters to ensure stable learning under limited data conditions. Results show that adaptive optimisers, especially Adam and Adamax, outperformed others. Adam achieved the highest accuracy (95.7%), while Adamax offered a better balance between accuracy and training efficiency. In contrast, Adadelta showed the poorest performance. Despite class imbalance and the absence of core data, the Adamax-optimised model achieved strong F1-scores across all facies. This study demonstrates the feasibility of optimised 1D-CNNs for lithofacies classification in data, and resource, constrained environments and highlights the importance of optimiser choice. Future work should include more diverse geological datasets and integrate core or seismic data for improved validation and broader applicability.