One of the key objectives in geophysics is to characterize the subsurface through the process of analyzing and interpreting geophysical field data that are typically acquired at the surface. Data-driven deep learning methods have enormous potential for accelerating and simplifying the process but also face many challenges, including poor generalizability, weak interpretability, and physical inconsistency. We present three strategies for imposing domain knowledge constraints on deep neural networks (DNNs) to help address these challenges. The first strategy is to integrate constraints into data by generating synthetic training datasets through geological and geophysical forward modeling and properly encoding prior knowledge as part of the input fed into the DNNs. The second strategy is to design nontrainable custom layers of physical operators and preconditioners in the DNN architecture to modify or shape feature maps calculated within the network to make them consistent with the prior knowledge. The final strategy is to implement prior geological information and geophysical laws as regularization terms in loss functions for training the DNNs. We discuss the implementation of these strategies in detail and demonstrate their effectiveness by applying them to geophysical data processing, imaging, interpretation, and subsurface model building.

Although computer science has seen remarkable advancements in foundation models, they remain underexplored in geoscience. Addressing this gap, we introduce a workflow to develop geophysical foundation models, such as data preparation, model pretraining, and adaption to downstream tasks. From 192 globally collected 3D seismic volumes, we create a carefully curated data set of 2,286,422 2D seismic images. To fully use these unlabeled images, we use self-supervised learning to pretrain a transformer-based seismic foundation model for producing all-purpose seismic features that work across various tasks and surveys. Through experiments on seismic facies classification, geobody identification, interpolation, denoising, and inversion, our pretrained model demonstrates versatility, generalization, scalability, and superior performance compared with baseline models. In conclusion, we provide a foundation model and vast data set to advance artificial intelligence (AI) in geophysics, addressing the challenges (poor generalization, a lack of labels, and repetitive training for task-specific models) of applying AI to geophysics and paving the way for future innovations in geoscience.

Earthquake monitoring is vital for understanding the physics of earthquakes and assessing seismic hazards. A standard monitoring workflow includes the interrelated and interdependent tasks of phase picking, association, and location. Although deep learning methods have been successfully applied to earthquake monitoring, they mostly address the tasks separately and ignore the geographic relationships among stations. Here, we propose a graph neural network that operates directly on multi-station seismic data and achieves simultaneous phase picking, association, and location. Particularly, the inter-station and inter-task physical relationships are informed in the network architecture to promote accuracy, interpretability, and physical consistency among cross-station and cross-task predictions. When applied to data from the Ridgecrest region and Japan, this method showed superior performance over previous deep learning-based phase-picking and localization methods. Overall, our study provides a prototype self-consistent all-in-one system of simultaneous seismic phase picking, association, and location, which has the potential for next-generation automated earthquake monitoring.

In seismology, while training a specific deep learning model for each task is common, it often faces challenges such as the scarcity of labeled data and limited regional generalization. Addressing these issues, we introduce SeisCLIP: a foundation model for seismology, leveraging contrastive learning during pretraining on multimodal data of seismic waveform spectra and the corresponding local and global event information. SeisCLIP consists of a transformer-based spectrum encoder and an multilayer perceptron (MLP)-based information encoder that are jointly pre-trained on massive data. During pretraining, contrastive learning aims to enhance representations by training two encoders to bring the corresponding waveform spectra and event information closer in the feature space, while distancing uncorrelated pairs. Remarkably, the pre-trained spectrum encoder offers versatile features, enabling its application across diverse tasks and regions. Thus, it requires only modest datasets for fine-tuning to specific downstream tasks. Our evaluations demonstrate SeisCLIP’s superior performance over baseline methods in tasks like event classification, localization, and focal mechanism analysis, even when using distinct datasets from various regions. In essence, SeisCLIP emerges as a promising foundational model for seismology, potentially revolutionizing foundation-model-based research in the domain.

Fault interpretation tasks become more and more difficult as the complexity of seismic exploration increases, especially for ultra-deep seismic data. Recently, numerous researchers have utilized automatic interpretation techniques based on deep learning to improve the efficiency and accuracy of fault prediction. Although deep learning methods have powerful data information processing capabilities, the applicability of deep neural networks may still be limited by the range of learned information. Therefore, we develop a new technique called structural data augmentation to enhance the diversity of the datasets. Concretely, we utilize different geological structure theories to incorporate virtual folds and faults in the field seismic data to improve the diversity and generalization ability of the training datasets. To cope with the multi-stage and multi-scale complex structures developed in ultra-deep strata, the proposed augmentation workflow increases data diversity by generating various virtual structures containing multi-scale folds, listric faults, oblique-slip displacement fields, and multi-directional fault drags. Tests on the field seismic data show that our method not only outperforms conventional seismic attributes but also has advantages over other machine learning methods.