Seismic Reflectivity Inversion Using a Semi-Supervised Learning Approach

Seismic reflectivity inversion is a critical component in subsurface characterization, playing an indispensable role in resource exploration and reservoir studies (Tarantola, 1984). Conventional inversion methods have traditionally relied on optimization-based techniques, which often necessitate prior low-frequency models, and are not only susceptible to noise but also prone to introducing biases due to their dependence on initial assumptions. In recent years, deep learning, using deep neural networks (DNNs), has emerged as a competitive approach for seismic inversion, capable of capturing complex relationships within data (Valentine & Trampert, 2012). However, supervised deep learning models are heavily dependent on the availability of labeled data, which is not always abundant or easily attainable (Alaudah & AlRegib, 2019). Furthermore, many existing models focus on predicting single points in reflectivity series, not adequately accounting for broader contexts within the seismic data (Wu et al., 2019). In this study, we present a novel semi-supervised deep learning approach for seismic reflectivity inversion that leverages both labeled and unlabeled data to produce high quality reflectivity estimates. Our model employs a sequence-to-sequence (Sutskever et al., 2014) architecture, operating directly on raw seismic data, and is capable of capturing broader contexts, thus ensuring the continuity and complexity of the reflectivity series. Additionally, it minimizes the reliance on prior low-frequency models, enhancing the model's generalization capabilities. We present a comprehensive evaluation of the proposed method through a series of experiments on synthetic datasets, contrasting its performance against conventional seismic inversion and supervised DNN methods. Moreover, we discuss the key advantages, potential limitations, and avenues for future research in the application of semi-supervised deep learning in seismic inversion.