Automated seismic fault mapping using convolutional neural networks for the Berenice Field, Western Desert, Egypt

Accurate fault interpretation remains a cornerstone of hydrocarbon exploration, directly influencing structural modeling, reservoir compartmentalization, and well placement decisions. However, conventional manual fault picking is time-intensive and prone to interpreter bias, especially in structurally complex and noisy seismic datasets such as those from Egypt's Western Desert. This study demonstrates the implementation and performance of Schlumberger's user-trained Convolutional Neural Network (CNN) based Fault Prediction Module within the Petrel E&P platform for automated fault detection in the Berenice Field, Faghur basin. The applied workflow integrates machine learning-based denoising, sparse interpreter-guided fault labeling, 3D probability cube generation, and fault extraction using thresholded planarity and azimuth attributes. Only three manually labeled seismic lines (< 1% of total volume) were sufficient for model training. The CNN model achieved a Dice coefficient exceeding 0.85 and successfully delineated 65% of the manually mapped faults, while also detecting additional previously unmapped subtle discontinuities. The ML-assisted workflow reduced interpretation time by over 80% compared to traditional methods, maintaining high geological coherence and consistency. These results confirm that machine learning-driven fault detection effectively enhances interpretational accuracy and efficiency in tectonically complex domains. The successful application to the Berenice Field underscores the potential of CNN-based fault prediction as a transformative tool for structural interpretation across Egypt's Western Desert and similar intracratonic settings worldwide.