"Deep Learning Application for Seismic Fault Interpretation Based on 3D Convolutional Neural Networks and ESSAttn Attention Mechanism"

This paper proposes a deep learning model based on 3D Convolutional Neural Networks (CNN) and a custom attention mechanism (ESSAttn) for seismic fault interpretation from 3D seismic data. The model combines the advantages of self-attention mechanisms and convolutional neural networks to enhance the ability to capture and represent features in three-dimensional seismic data. The core innovation of the model lies in the introduction of the ESSAttn layer, which applies a non-traditional normalization process to the input feature queries, keys, and values, thereby strengthening the relationships between features, especially in high-dimensional seismic data. Unlike traditional attention mechanisms, the ESSAttn layer normalizes feature vectors by squaring them and integrates features across depth, width, height, and channel dimensions, significantly improving the effectiveness of attention computation.

The model's role in seismic fault interpretation is reflected in several aspects. First, the 3D convolutional layers automatically extract spatial features from seismic data, accurately capturing the location and shape of faults. Second, the ESSAttn layer enhances critical region features and focuses attention on important areas such as fault zones, reducing the interference from background noise and significantly improving fault detection accuracy. Finally, by using a weighted binary cross-entropy loss function, the model can prioritize fault regions when handling imbalanced data, improving sensitivity to weak fault signals.

The network architecture consists of three main parts: encoding, attention enhancement, and decoding. Initially, two 3D convolutional layers and max-pooling layers are used for feature extraction and down-sampling, followed by the ESSAttn layer to enhance the extracted features. The decoding part restores spatial resolution through upsampling and convolution layers, ultimately outputting the fault prediction results. The model is trained using the Adam optimizer, with a learning rate set to 1e-4.

Experimental results show that the model performs well in seismic fault interpretation tasks, effectively extracting and enhancing fault-related features. It is particularly suitable for automatic fault identification and localization in complex geological environments. The model's automation of feature extraction and enhancement reduces manual intervention, increases analysis efficiency, and demonstrates strong adaptability to large-scale 3D seismic datasets. Furthermore, the model architecture was visualized and saved using visualization tools for easier analysis and presentation.

Keywords: 3D Convolutional Neural Networks, ESSAttn, Attention Mechanism, Fault Interpretation, Weighted Cross-Entropy, 3D Seismic Data, Deep Learning