Velocity model inversion is one of the most important tasks in seismic exploration. Full-waveform inversion (FWI) can obtain the highest resolution in traditional velocity inversion methods, but it heavily depends on initial models and is computationally expensive. In recent years, a large number of deep-learning (DL)-based velocity model inversion methods have been proposed. One critical component in those DL-based methods is a large training set containing different velocity models. We have developed a method to construct a realistic structural model for the DL network. Our compressional-wave velocity model building method for creating dense-layer/fault/salt body models can automatically construct a large number of models without much human effort, which is very meaningful for DL networks. Moreover, to improve the inversion result on these realistic structural models, instead of only using the common-shot gather, we also extract features from the common-receiver gather as well. Through a large number of realistic structural models, reasonable data acquisition methods, and appropriate network setups, a more generalized result can be obtained through our proposed inversion framework, which has been demonstrated to be effective on the independent testing data set. The results of dense-layer models, fault models, and salt body models that we compared and analyzed demonstrate the reliability of our method and also provide practical guidelines for choosing optimal inversion strategies in realistic situations.

Seismic velocity inversion plays a vital role in various applied seismology processes. A series of deep learning methods have been developed that rely purely on manually provided labels for supervision; however, their performances depend heavily on using large training data sets with corresponding velocity models. Because no physical laws are used in the training phase, it is usually challenging to generalize trained neural networks to a new data domain. To mitigate these issues, we have embedded a seismic forward modeling step at the end of a network to remap the inversion result back to seismic data and thus train the neural network through self-supervised loss, i.e., the misfit between the network input and output. As a result, we eliminate the need for many labeled velocity models, and physical laws are introduced when back-propagating gradients through the seismic forward modeling step. We verify the effectiveness of our approach through comprehensive experiments on synthetic data sets, where self-supervised learning outperforms the fully supervised approach, which accesses much more labeled data. The superior performance is even more significant when compared with a new data domain that has velocity models with faults and more geologic layers. Finally, in case of unknown and more complex data types, we develop a network-constrained full-waveform inversion (FWI) method. This method refines the initial prediction of the network by iteratively optimizing network parameters other than the velocity model, as found with the conventional FWI method, and demonstrates clear advantages in terms of interface and velocity accuracy. With these measures (self-supervised learning and network-constrained FWI), our physics-driven self-supervised learning system successfully mitigates issues such as the dependence on large labeled data sets, the absence of physical laws, and the difficulty in adapting to new data domains.

Training a deep learning inversion network usually requires hundreds of thousands of complex velocity models, which is labor-intensive and expensive to acquire. In this work, we develop a new framework to automatically generate various velocity models with common geological structures, such as folding layers, faults and salt bodies. There are three main modules in the proposed framework. The first module generates a folded model with a given number of layers; the other two modules can add faults and salt bodies onto the folded model to form a fault or salt model, respectively. To best simulate the shape of subsurface geological structures while ensuring a good application effect in deep learning inversion, we generate the structural model based on composition of several basic functions with recommended parameter ranges. Then the generated structural model will be assigned with velocity values based on assumptions of underground seismic velocity distribution. To investigate the application effect of the generated 3D models, we conduct a deep learning inversion test. Since currently there is no 3D deep learning inversion algorithms available, the latest 2D inversion network called SeisInvNet is used to test the feasibility of the randomly generated velocity models. Through the experiments, we can see that the 2D inversion results are consistent with the true models from the aspects of velocity values and geological structure shapes, which demonstrates the rationality of the designed 3D models. In the end, we further discuss the feasibility of applying the proposed 3D model dataset to train a 3D inversion network. This work paves the way for the development of 3D deep learning inversion methods.

Random noise attenuation is an essential step to improve the signal-to-noise ratio (SNR) of seismic data. Deep learning for seismic data denoising is dominated by supervised methods that require noise-free data as training targets. It is usually time-consuming and laborious to obtain such clean seismic data, and the effectiveness of the noise attenuation is difficult to be guaranteed. Therefore, we propose a novel unsupervised learning method that learns from noisy data. The method is based on two salient features of seismic data: 1) valid signals of adjacent seismic traces that are spatially correlated and 2) random noise that is spatially independent and unpredictable. An end-to-end deep convolutional neural network (CNN) was constructed to solve the denoising task. Adjacent traces of seismic data, which contain similar seismic phases and interface features, were used as the inputs and labels of the training set. The mapping of spatial correlation can be learned by the CNN so that valid signals from raw seismic data are predicted, while random noise is suppressed for unpredictability. Synthetic and field data were applied to the proposed CNN denoising model. The experimental results demonstrate the effectiveness of random noise attenuation while preserving amplitude compared with two commonly used state-of-the-art denoising methods.

Seismic forward prospecting is essential because it can identify the velocity distribution in front of the tunnel face and provide guidance for safe excavation activities. We have developed a convolutional neural network (CNN)-based method to invert forward-prospecting data recorded in tunnels for accurate and rapid estimation of seismic velocity distribution. Targeting the unusual seismic acquisition setup in tunnels, we design two separate encoders to extract features from observation data recorded on both tunnel sidewalls. Subsequently, these features are concatenated to a decoder for velocity prediction. Considering the various acquisition setups used in different tunneling projects, the deep-learning inversion network must be flexible in terms of the seismic source/receiver positions for practical application. We have generated two auxiliary feature maps that can be used to feed acquisition information to our network. Our network, acquisition adaptive CNN (A2-CNN), can be trained by defining the loss function based on the L2-norm and multiscale structural similarity. Compared with traditional CNNs, our method has superior performance on data sets with fixed and random acquisition setups and also demonstrates certain robustness when handling synthetic data with field noise. Finally, we test how the network performs when feeding the modified acquisition setup information. It turns out that the inversion result will demonstrate a shift when the provided acquisition setup information shift, which verified the validity of the network and its use of acquisition information.

The high-resolution waveform inversion for seismic velocities is gaining increasing interest as we start to deal with complex structures. Although full waveform inversion has been used for several years, obtaining high-resolution velocity models still presents many obstacles, such as the high computational cost and the limited band width of the data. Thus, we propose a deep learning-based algorithm to build high-resolution velocity models using low-resolution velocity models, migration images, and well-log velocities as inputs. The well information, specifically, helps enhance the resolution with ground truth information, especially around the well. These three inputs are fed to an improved neural network, a variant of U-Net, as three channels to predict the corresponding true velocity models, which serve as labels in the training. The incorporation of well velocities from several locations is crucial for improving the resolution of the output model. Numerical experiments on complex models demonstrate the robust performance of this network and the crucial role that well information plays, especially in generalizing the approach to models that differ from the trained ones and achieving superior performance compared to full waveform inversion.