Q-factor and Reflectivity Estimation Using Semi-Supervised Hybrid Deep Neural Networks

The quality factor and reflectivity are important for reservoir identification. Traditional Q estimation methods are susceptible to noise, and the conversion of equivalent Q to interval Q can easily accumulate errors. Additionally, traditional reflectivity inversion methods do not adequately consider the nonstationary variations of wavelets during subsurface propagation, leading to significantly reduced inversion accuracy for deep reflection interfaces. Deep learning, with its powerful nonlinear modeling capability, provides an effective solution to such inverse problems. Generally, deep learning approaches typically demand large quantities of labeled data, whereas the inversion problem under consideration suffers from severe label scarcity. This study proposes a Semi-Supervised Hybrid Deep Neural Network (HDNN) for interval Q and reflectivity inversion. The proposed HDNN architecture effectively integrates Convolutional Neural Network (CNN), Transformer, and Gated Recurrent Unit (GRU). Furthermore, Margrave’s nonstationary convolutional model is employed to synthetic seismogram. Our method utilizes both limited labeled and abundant unlabeled seismic data within a joint training framework. The network is optimized using a composite loss function that combines supervised and unsupervised terms, achieving semi-supervised learning. Applications to both synthetic and field data demonstrate that our proposed semi-supervised hybrid neural network can reliably estimate interval Q and reflectivity.