Transformer Model for Fault Detection from Brazilian Pre-salt Seismic Data

• Letícia Bomfim UNICAMP / CEPETRO
• Oton Cunha CEPETRO
• Michelle Kuroda CEPETRO
• Alexandre Vidal UNICAMP / CEPETRO
• Helio Pedrini UNICAMP

Resumo

Carbonate reservoirs are known for their heterogeneity, which poses challenges in interpreting and defining geological models. The Brazilian reservoirs are formed mostly in highly faulted and fractured carbonate rocks, which can increase hydrocarbon transport and storage capacity. Strategies that permit the identification of these structures allow the optimization in the exploration of a reservoir. To fulfill this task, machine learning models have been able to provide an understanding of these environments through the use of data obtained by seismic methods. The use of convolutional neural networks (CNNs) has shown to be able to provide excellent abstractions in the field of semantic segmentation, including its use in seismic data. However, due to the highly heterogeneous formation of this type of data, the work of extracting information from these images remains challenging. From this, we investigate the potential of using Transformer models in this geological context focusing on the faults identification. As a technique to analyze this type of architecture, we use the TransUNet network, which combines the power of CNNs with the innovation brought by Transformers in a hybrid model of deep learning. To evaluate its performance, we made use of conventional CNNs to compare the results achieved. The results show that TransUNet outperforms conventional CNNs, achieving a Dice metric value of 88.34%, compared to 85.99% for U-Net, 83.41% for U-Net++, and 83.31% for SegNet, being also able to identify small structures beyond what is indicated in our target.