Seismic Impedance Inversion Using a Joint Deep Learning Model Based on Convolutional Neural Network and Transformer

Abstract

Seismic impedance is an important factor in characterizing reservoirs, so accurate seismic impedance inversion is significant in seismic exploration. However, achieving high-resolution impedance inversion has remained a complex problem due to challenges related to the unknown seismic wavelet and the frequency band limitations of the observed data. In recent years, deep learning methods such as convolutional neural network (CNN) have been successfully applied to the field of seismic impedance inversion, which can obtain higher resolution results compared with traditional inversion methods. However, limited by the size of the local receptive field, CNN is not conducive to extracting global information. In contrast, a transformer can efficiently extract long-range dependencies but relies entirely on the self-attention mechanism to compute correlations between data, which requires a lot of training data. Therefore, this article proposes a joint deep learning model based on CNN and a transformer for impedance inversion. Among them, CNN and transformer are used to learn local and global information in the data, respectively, and feature fusion through residual connection, which can improve the feature representation capability of neural networks. In addition, we train a CNN-based forward operator that can introduce information from unlabeled data into the network training to enhance the network's generalization ability and improve the stability of the inversion. Experimental results in the SEAM model and field data show that the method can predict impedance effectively and with better accuracy than classical constrained sparse spike inversion and conventional deep learning methods.