plane-wave destruction

This study addresses the limited resolution of seismic velocity models caused by bandwidth constraints in seismic data. To overcome this challenge, the authors propose a diffusion-guided high-resolution velocity modeling framework that innovatively embeds local geological dip information into the diffusion prior. By jointly generating velocity and dip priors, the method propagates sparse well data along structural orientations and integrates plane-wave PDE regularization, structure-preconditioned inversion, and measurement-guided DDIM posterior sampling to enable efficient, structurally constrained reconstruction. Experiments on both the Volve synthetic model and real-field Viking Graben data demonstrate that the proposed approach significantly enhances structural continuity, lateral consistency, and geological plausibility of the reconstructed velocity models while maintaining high computational efficiency.

To address the band-limited and geologically distorted reconstructions in full-waveform inversion (FWI) caused by insufficient seismic observations, this paper proposes a conditional diffusion model-driven FWI framework integrating multiple geological priors. For the first time, well-log curves and geological facies labels are jointly embedded as multimodal conditioning inputs into a diffusion model, coupled with classifier-free guidance to enforce joint geophysical–geological constraints—thereby overcoming the limitations of conventional single-velocity-prior regularization. Evaluated on the OpenFWI benchmark and real marine seismic data, the method significantly improves both inversion accuracy and geological plausibility. It outperforms standard FWI and unconditional diffusion-regularized FWI, demonstrating superior robustness under data-scarce conditions. This work establishes a novel paradigm for high-fidelity subsurface imaging when seismic coverage is limited.

Marine towed-streamer seismic data often suffer from missing near-offset traces, which severely compromises critical processing tasks such as multiple suppression, velocity analysis, and full-waveform inversion. This work proposes the first self-supervised diffusion framework to progressively reconstruct missing near-offset traces from available far-offset data without requiring ground-truth labels. By integrating a conditional diffusion model with a trace-wise recursive extrapolation strategy, the method leverages sliding overlapping patch extraction and ensemble sampling to achieve superior performance over conventional parabolic Radon transform approaches on both synthetic and field datasets. The framework not only faithfully recovers amplitude-versus-offset (AVO) characteristics but also provides uncertainty quantification, effectively highlighting regions where extrapolation is particularly challenging.

This study addresses the challenge of uncertainty quantification in geophysical probabilistic inversion by systematically introducing Flow Matching—a generative artificial intelligence technique—into full-waveform inversion for the first time. The method constructs a continuous transformation path from the prior to the posterior distribution of subsurface velocity models, parameterized by deep neural networks. Within a Bayesian framework, it leverages continuous normalizing flows to enable efficient posterior sampling. The approach demonstrates stable and accurate uncertainty estimation on both a 2D synthetic model and the complex OpenFWI benchmark dataset, significantly extending the applicability of Flow Matching to scientific computing and inverse problem solving.

To address insufficient accuracy in high-frequency (≤50 Hz) seismic waveform prediction and poor generalization under sparse data, this paper introduces the first conditional generative model based on denoising diffusion probabilistic modeling. It synthesizes waveforms efficiently in the latent space of a spectrogram-based autoencoder, conditioned on magnitude, fault distance, site conditions, and fault type. Innovatively integrating diffusion modeling with time–frequency representation, the method significantly enhances extrapolation capability in data-sparse regimes and supports inversion of arbitrary scalar ground-motion parameters. Experiments demonstrate that generated waveforms faithfully reproduce both the median trends and variability of real observations, achieving state-of-the-art performance on seismic metrics (e.g., PGA, PSA) and image-quality metrics (e.g., FID, LPIPS). The open-source implementation establishes an interpretable and scalable paradigm for seismic hazard analysis and earthquake-resistant design.

Unsupervised 3D seismic horizon tracking often fails near faults: signal-driven methods offer high precision but poor robustness, while texture-driven approaches handle discontinuities yet rely on labels and suffer from limited local accuracy. This work proposes a self-supervised contrastive learning framework that integrates both signal and texture cues, introducing for the first time inter-trace flow derived from reflection dip estimation as a domain-specific prior. Positive sample pairs are constructed within high-confidence neighborhoods to propagate horizon identity consistently across faults. The method trains a texture-aware voxel embedding model by combining high-confidence region constraints with an optional fault mask. Evaluated on the public F3 dataset and synthetic data with faults, the approach achieves a mean absolute error (MAE) significantly better than unsupervised baselines and comparable to semi-supervised methods using only a single annotated slice.

This study addresses the challenges of weak model generalization, lack of physical consistency, and insufficient uncertainty quantification in detecting seismic and volcanic signals under non-stationary, high-noise conditions. To overcome these limitations, the work integrates physical priors with machine learning by embedding classical signal processing as an inductive bias, leveraging self-supervised and generative modeling to reduce reliance on labeled data, and introducing a more realistic cross-regional transfer evaluation protocol. The proposed approach significantly enhances model generalization, interpretability, and decision reliability in phase picking and anomaly detection tasks—particularly in unseen stations, time-varying noise environments, and evolving volcanic activity—thereby advancing the practical deployment of AI in geophysical monitoring.

This study addresses the challenge of covariance estimation in Small Baseline Subset (SBAS) interferometry caused by systematic, large-scale data gaps. Viewing two-dimensional partial observations as fragmented functional data, the authors introduce a novel “fragmented observation mechanism” framework from the perspective of functional data analysis. They propose a Laplacian-regularized matrix completion method that enables nonparametric covariance estimation without requiring assumptions of stationarity or isotropy. The approach demonstrates consistently low estimation errors across diverse covariance structures in simulations and is successfully applied to real SBAS-derived surface deformation data from the Phlegraean Fields. The method effectively recovers spatial dependence patterns, offering robust support for environmental risk monitoring.

This study addresses the challenge of strong surface-wave noise in land and vertical seismic profile (VSP) data, which often overlaps significantly with useful reflection signals, limiting the adaptability of conventional denoising methods or rendering them dependent on labeled training data. To overcome this, the authors propose a semantic-guided signal separation approach that, for the first time, leverages a large vision model (LVM)—requiring no task-specific training—to extract surface-wave semantic priors from visual representations of seismic gathers via text or image prompts. These priors are used to generate continuous soft masks, which are then integrated into a mask-constrained low-rank inversion framework solved via the ADMM algorithm for adaptive denoising. The method requires neither annotated data nor fine-tuning and consistently outperforms traditional filtering and implicit neural representation techniques on both synthetic and real VSP datasets, effectively attenuating surface waves while preserving reflection continuity and waveform fidelity.

This work addresses the challenge of generating geologically plausible and statistically consistent high-resolution velocity models from incomplete observations, such as sparse well logs and migrated seismic images. To this end, the authors propose SAGE, a multimodal generative framework that learns a surrogate posterior distribution of velocity models conditioned on both well data and seismic images. During inference, SAGE requires only the seismic image to produce full-resolution velocity fields, implicitly incorporating well information through the learned generative prior. SAGE is the first method capable of generating geologically realistic and statistically consistent velocity models using solely migrated seismic images. Validated on both synthetic and field datasets, it accurately captures complex subsurface structures and provides high-quality, diverse velocity realizations suitable for downstream inversion tasks.