我们介绍了为自动神经网络处理设计的Seismic Waveforms dataset for Automatic Neural-network processing (SWAN)，一个全面且标准化的基准，旨在推动数据驱动的地震信号处理。SWAN聚合了涵盖广泛地质结构、噪声条件、传播环境和采集几何的合成和真实地震波形，为训练高度通用的模型提供统一基础。利用该数据集，我们开发并评估了条件约束残差扩散模型，专注于核心地震处理任务中的缺失轨迹重建。大量实验表明，在SWAN上训练的扩散模型在异构测试场景中均取得最佳性能，优于领先的深度学习和物理基础基线，在合成基准和现场数据示例中均表现优异。结果突显了SWAN作为可扩展训练语料库和严格评估框架的价值，并展示了基于扩散架构的强大力量，用于稳健、通用的地震数据处理。
英文摘要：
We introduce the Seismic Waveforms dataset for Automatic Neural-network processing (SWAN), a comprehensive and standardized benchmark designed to advance data-driven seismic signal processing. SWAN aggregates diverse synthetic and real seismic waveforms spanning a wide range of geological structures, noise conditions, propagation environments, and acquisition geometries, providing a unified foundation for training highly generalizable models. Leveraging this dataset, we develop and evaluate a conditionally constrained residual diffusion model for core seismic processing tasks, focusing on missing-trace reconstruction. Extensive experiments demonstrate that diffusion models trained on SWAN achieve state-of-the-art performance across heterogeneous testing scenarios, outperforming leading deep-learning and physics-based baselines on both synthetic benchmarks and field data examples. The results highlight SWAN's value as both a scalable training corpus and a rigorous evaluation framework, and illustrate the strong potential of diffusion-based architectures for robust, generalizable seismic data processing.
The idea of curvature analysis has been widely used in subsurface structure interpretation from three-dimensional (3D) seismic data (e.g., fault/fracture detection and geomorphology delineation) by measuring the lateral changes in the geometry of seismic events. However, such geometric curvature utilizes only the kinematic information (two-way traveltime) of the available seismic signals. While analyzing the dynamic information (waveform), the traditional approaches (e.g., complex trace analysis) are often trace-wise and thereby fail to take into account the seismic reflector continuity and de…
We present a technique for reconstructing subsurface velocity model changes from time-lapse seismic survey data using full-waveform inversion (FWI). The technique is based on simultaneously inverting multiple survey vintages, with model difference regularization using the total variation (TV) seminorm. We compare the new TV-regularized time-lapse FWI with the $L_2$-regularized joint inversion proposed in our earlier work, using synthetic data sets that exhibit survey repeatability issues. The results demonstrate clear advantages of the proposed TV-regularized joint inversion over alternatives …
Waveform-Based Probabilistic Seismic Hazard Analysis Using Ground-Motion Generative Models
National Research Institute for Resilience Engineering Research Center Earth Science and Disaster Resilience Graduate School of Engineering（国立研究开发法人防灾科学技术研究所地球科学与灾害韧性研究生院）；The University of Tokyo（东京大学）
I train a Generative Adversarial Network to produce realistic seismic wave speed models. I integrate the generator network into seismic Full-Waveform Inversion to reduce the number of model parameters and restrict the inverted models to only those that are plausible. Applying the method to a 2D section of the SEAM model, I demonstrate that it can produce more plausible results than conventional Full-Waveform Inversion.
Convergent representations of elastic-wave structure emerge across deliberately distinct seismic training routes
Institute of Geophysics, China Earthquake Administration（中国地震局地球物理研究所）；University of Chinese Academy of Sciences（中国科学院大学）；Laboratory of Seismology and Physics of Earth’s Interior, School of Earth and Space Sciences, University of Science and Technology of China（中国科学技术大学地球和空间科学学院地球内部地震学与物理学实验室）；Institute of Advanced Technology, University of Science and Technology of China（中国科学技术大学先进技术研究院）
Using spectral-element and adjoint methods, we investigate body waves interacting with the Earth's most dramatic interface, the core-mantle boundary (CMB). Intermediate-to-high frequency seismograms are computed incorporating topography models. We analyse the sensitivity of many seismic phases interacting with the interface. The study aims at showing effects of CMB structure on synthetics and highlights difficulties of imaging this region due to strong trade-off between velocity variations and topography. Synthetic waveforms computed at dominant periods of 6-18 seconds are used in order to obs…
Derivatives of the displacement tensor with respect to the independent model parameters of the subsurface, also called Fréchet derivatives (or sensitivity kernels), are a key ingredient for seismic full-waveform inversion with a local-search optimization algorithm. They provide a quantitative measure of the expected changes in the seismograms due to perturbations of the subsurface model parameters for a given survey geometry. Since 2.5-D wavefield modeling involves a real point source in a 2-D geological model with 3D (spherical) wave properties, it yields synthetic data much closer to the act…
Seismic full-waveform inversion (FWI) uses full seismic records to estimate subsurface velocity structure. This requires a highly nonlinear and nonunique inverse problem to be solved, and Bayesian methods have been used to quantify uncertainties in the solution. Variational Bayesian inference uses optimization to provide solutions efficiently. The method has been applied to solve a transmission FWI problem using data generated by known earthquake-like sources, with strong prior information imposed on the velocity. Unfortunately such prior information about velocity structure and earthquake sou…
Traditional seismic envelope inversion takes use of a nonlinear misfit functional which relates the envelope of seismogram to the observed wavefield records, and then derive the sensitivity kernel of envelope to velocity through the use of waveform Frechet derivative by linearizing the nonlinear functional. We know that the waveform Frechet derivative is based on the Born approximation of wave scattering and can only work for the case of weak scattering. Therefore, the traditional envelope inversion using waveform Frechet derivative has severe limitation in the case of strong-scattering for la…
Low-magnitude seismic activity between the Kamchatka July 20 and July 29, 2025, earthquakes. Spatio-temporal evolution recovered using waveform cross-correlation
We have implemented an extension for the observational seismology obspy software package to provide a streamlined tool tailored to the processing of seismic signals from non-earthquake sources, in particular those from deforming systems such as glaciers and landslides. This seismic attributes library provides functionality to: (1) download and/or pre-process seismic waveform data; (2) detect and catalogue seismic events using multi-component signals from one or more seismometers; and (3) calculate characteristics ('attributes'/'features') of the identified events. The workflow is controlled by…
I demonstrate that the conventional seismic full-waveform inversion algorithm can be constructed as a recurrent neural network and so implemented using deep learning software such as TensorFlow. Applying another deep learning concept, the Adam optimizer with minibatches of data, produces quicker convergence toward the true wave speed model on a 2D dataset than Stochastic Gradient Descent and than the L-BFGS-B optimizer with the cost function and gradient computed using the entire training dataset. I also show that the cost function gradient calculation using reverse-mode automatic differentiat…
University of Basel（巴塞尔大学）；Université Grenoble Alpes（格勒诺布尔阿尔卑斯大学）；Université Savoie Mont Blanc（萨瓦勃朗峰大学）；CNRS（法国国家科学研究中心）；Université Gustave Eiffel（埃菲尔大学）；Sapienza University of Rome（罗马大学）；Pennsylvania State University（宾夕法尼亚州立大学）；Rice University（莱斯大学）
We consider the high-resolution seismic imaging method called full-waveform inversion (FWI). FWI is a data fitting method aimed at inverting for subsurface mechanical parameters. Despite the large adoption of FWI by the academic and industrial communities, and many successful results, FWI still suffers from severe limitations. From a mathematical standpoint, FWI is a large scale PDE-constrained optimization problem. The misfit function that is used, which measures the discrepancy between observed seismic data and data calculated through the solution of a wave propagation problem, is non-convex…
We propose and test the Direct Waveform Inversion (DWI) scheme to simultaneously invert for layered velocity and density profiles, using reflection seismic waveforms recorded on the surface. The recorded data include primary reflections and interbed multiples. DWI is implemented in the time-space domain. By explicitly enforcing the wavefield time-space causality, DWI can recursively determine the subsurface seismic structure in a local fashion for both sharp interfaces as well as the properties of the layers, from shallow to deep depths. DWI is different from the layer stripping methods in the…
In seismic waveform inversion, the reconstruction of the subsurface properties is usually carried out using approximative wave propagation models to ensure computational efficiency. The viscoelastic nature of the subsurface is often unaccounted for, and two popular approximations--the acoustic and linearized Born inversion--are widely used. This leads to reconstruction errors since the approximations ignore realistic (physical) aspects of seismic wave propagation in the heterogeneous earth. In this study, we show that the Bayesian approximation error approach can be used to partially recover f…
Active faults release tectonic stress imposed by plate motion through a spectrum of slip modes, from slow, aseismic slip, to dynamic, seismic events. Slow earthquakes are often associated with tectonic tremor, non-impulsive signals that can easily be buried in seismic noise and go undetected. We present a new methodology aimed at improving the detection and location of tremors hidden within seismic noise. After identifying tremors with a classic convolutional neural network, we rely on neural network attribution to extract core tremor signatures and denoise input waveforms. We then use these c…
Prediction of the Kamchatka July 29, 2025, earthquake by the evolution of low-magnitude seismicity recovered using waveform cross-correlation at IMS seismic arrays
In the last ten years, full-waveform inversion has emerged as a robust and efficient high-resolution velocity model-building tool for seismic imaging, with the unique ability to recover complex subsurface structures. Originally based on a data fitting process using a least-squares cost function, it suffered from high sensitivity to cycle-skipping and was therefore of poor efficiency in handling large time shifts between observed and modelled seismic events. To tackle this problem, a common practice is to start the inversion using the low temporal frequencies of the data and selecting diving wa…
Geophysical models usually contain both sharp interfaces and smooth variations, and it is difficult to accurately account for both of these two types of medium parameter variations using conventional full-waveform inversion methods. In addition, sparse geometry, noisy data and source encoding usually lead to strong inversion artifacts. We develop a novel full-waveform inversion method for acoustic and elastic waves using a total generalized p-variation regularization scheme to address these challenging problems. We decompose the full-waveform inversion into two subproblems and solve these two …
In this work, we propose a full-waveform technique for the spatial reconstruction and characterization of (micro-) seismic events via joint source location and moment tensor inversion. The approach is formulated in the frequency domain, and it allows for the simultaneous inversion of multiple point-like events. In the core of the proposed methodology is a grid search for the source locations that encapsulates the optimality condition on the respective moment tensors. The developments cater for compactly supported elastic bodies in $\mathbb{R}^2$; however our framework is directly extendable to…
Precise real time estimates of earthquake magnitude and location are essential for early warning and rapid response. While recently multiple deep learning approaches for fast assessment of earthquakes have been proposed, they usually rely on either seismic records from a single station or from a fixed set of seismic stations. Here we introduce a new model for real-time magnitude and location estimation using the attention based transformer networks. Our approach incorporates waveforms from a dynamically varying set of stations and outperforms deep learning baselines in both magnitude and locat…
Recovery of low-magnitude seismic events before the July 29, 2025, Kamchatka megathrust earthquake using waveform cross-correlation enhanced by the addition of stochastic noise
Non-invasive subsurface imaging using full waveform inversion (FWI) has the potential to fundamentally change engineering site characterization by enabling the recovery of high resolution 2D/3D maps of subsurface stiffness. Yet, the accuracy of FWI remains quite sensitive to the choice of the initial starting model due to the complexity and non-uniqueness of the inverse problem. In response, we present the novel application of convolutional neural networks (CNNs) to transform an experimental seismic wavefield acquired using a linear array of surface sensors directly into a robust starting mode…
We present a new seismic inversion method that uses deep learning (DL) features for the subsurface velocity model estimation. The DL feature is a low-dimensional representation of the high-dimensional seismic data, which is automatically generated by a convolutional autoencoder (CAE) and preserved in the latent space. The low-dimensional DL feature contains the key information of the input seismic data. Therefore, instead of directly comparing the waveform differences between the observed and predicted data, such as full-waveform inversion (FWI). We measure their DL feature differences in the …
Los Alamos National Laboratory（洛斯阿拉莫斯国家实验室）；The Pennsylvania State University（宾夕法尼亚州立大学）；Michigan State University（密歇根州立大学）；Microsoft Research（微软研究院）
This paper investigates unsupervised learning of Full-Waveform Inversion (FWI), which has been widely used in geophysics to estimate subsurface velocity maps from seismic data. This problem is mathematically formulated by a second order partial differential equation (PDE), but is hard to solve. Moreover, acquiring velocity map is extremely expensive, making it impractical to scale up a supervised approach to train the mapping from seismic data to velocity maps with convolutional neural networks (CNN). We address these difficulties by integrating PDE and CNN in a loop, thus shifting the paradig…
This paper presents a novel numerical method for the Newton seismic full-waveform inversion (FWI). The method is based on the full-space approach, where the state, adjoint state, and control variables are optimized simultaneously. Each Newton step is formulated as a PDE-constrained optimization problem, which is cast in the form of the Karush-Kuhn-Tucker (KKT) system of linear algebraic equitations. The KKT system is solved inexactly with a preconditioned Krylov solver. We introduced two preconditioners: the one based on the block-triangular factorization and its variant with an inexact block …
Relative relocation methods are commonly used to precisely relocate earthquake clusters consisting of similar waveforms. Repeating waveforms are often recorded at volcanoes, where, however, the crust structure is expected to contain strong heterogeneities and therefore the 1D velocity model assumption that is made in most location strategies is not likely to describe reality. A peculiar cluster of repeating low-frequency seismic events was recorded on the south flank of Katla volcano (Iceland) from 2011. As the hypocentres are located at the rim of the glacier, the seismicity may be due to vol…
Detecting weak seismic events from noisy sensors is a difficult perceptual task. We formulate this task as Bayesian inference and propose a generative model of seismic events and signals across a network of spatially distributed stations. Our system, SIGVISA, is the first to directly model seismic waveforms, allowing it to incorporate a rich representation of the physics underlying the signal generation process. We use Gaussian processes over wavelet parameters to predict detailed waveform fluctuations based on historical events, while degrading smoothly to simple parametric envelopes in regio…
Full-waveform inversion (FWI) is a widely used technique in seismic processing to produce high resolution Earth models that fully explain the recorded seismic data. FWI is a local optimisation problem which aims to minimise in a least-squares sense the misfit between recorded and modelled data. The inversion process begins with a best-guess initial model which is iteratively improved using a sequence of linearised local inversions to solve a fully non-linear problem. Deep learning has gained widespread popularity in the new millennium. At the core of these tools are Neural Networks (NN), in pa…
We present an application of deep generative models in the context of partial-differential equation (PDE) constrained inverse problems. We combine a generative adversarial network (GAN) representing an a priori model that creates subsurface geological structures and their petrophysical properties, with the numerical solution of the PDE governing the propagation of acoustic waves within the earth's interior. We perform Bayesian inversion using an approximate Metropolis-adjusted Langevin algorithm (MALA) to sample from the posterior given seismic observations. Gradients with respect to the model…
Seismic-phase detection using multiple deep learning models for global and local representations of waveforms
Earthquake Research Institute, The University of Tokyo（东京大学地震研究所）；Graduate School of Information Science and Technology, The University of Tokyo（东京大学信息科学与技术研究生院）
