The existence of coal seams suppresses other useful information, especially the below-thin layers, and is unfavorable for delineating the target reservoirs beneath them. The matching pursuit (MP) based methods are commonly used for removing strong reflections caused by coal seams. They first decompose a seismic trace into several wavelets based on a user-defined wavelet dictionary and then separate the most similar wavelet with the coal seam. However, how to define a complete wavelet dictionary and how to maintain horizontal continuity are two unsolved issues. We propose a multi-scale attention deep learning (MSADL) model for separating and removing seismic strong reflections. First, we suggest a workflow to generate a synthetic data set for model training based on the characteristics of field data and well logs. Next, we build an MSADL model by integrating the discrete wavelet transform (DWT) and convolutional block attention module (CBAM) into the widely used Unet. After model training, we apply the well-trained MSADL model to 3D field data in the Sichuan Basin, China for the separation and removal of strong reflections and characterization of the beneath target thin layers.

Predicting the petrophysical properties of rocks from seismic data is challenging because the relationship between petrophysical properties and their seismic response is non-linear and multi-solution. A targeted framework is presented to enhance the effectiveness of petrophysical properties prediction based on deep learning in this study. We utilize statistical rock physics and geological method to tackle the challenge of the limited availability of labeled data. Moreover, in response to the issue of multiple solutions, we propose multi-task inversion neural network architecture for predicting petrophysical properties from multi-information guided by physical model. To evaluate the validity of the framework, we build a numerical model and carried out a quantitative analysis using 3-fold cross-validation and comparison of single trace predictive results. The framework is finally applied to a real work area of deep tight dolomite reservoir in Southwest China, demonstrating promising prospects for practical application.

The total organic carbon (TOC) is an important parameter for shale gas reservoir exploration. Currently, predicting TOC using seismic elastic properties is challenging and of great uncertainty. The inverse relationship, which acts as a bridge between TOC and elastic properties, is required to be established correctly. Machine learning especially for Random Forests (RF) provides a new potential. The RF-based supervised method is limited in the prediction of TOC because it requires large amounts of feature variables and is very onerous and experience-dependent to derive effective feature variables from real seismic data. To address this issue, we propose to use the extended elastic impedance to automatically generate 222 extended elastic properties as the feature variables for RF predictor training. In addition, the synthetic minority oversampling technique is used to overcome the problem of RF training with imbalanced samples. With the help of variable importance measures, the feature variables that are important for TOC prediction can be preferentially selected and the redundancy of the input data can be reduced. The RF predictor is finally trained well for TOC prediction. The method is applied to a real dataset acquired over a shale gas study area located in southwest China. Examples illustrate the role of extended variables on improving TOC prediction and increasing the generalization of RF in prediction of other petrophysical properties.