GlobalTomo: A global dataset for physics-ML seismic wavefield modeling and FWI

TL;DR

GlobalTomo addresses the computational bottlenecks of global seismic tomography by providing a first-of-its-kind 3D global synthetic dataset for forward modeling and full-waveform inversion, generated with AxiSEM3D and parameterized via spherical harmonics up to degree 8. The dataset comprises Acoustic, Elastic, and Real Earth tiers, offering wavefields and seismograms to support physics-informed machine learning and neural operator training. A suite of ML baselines (including DeepONet, gnot, and physics-informed variants) demonstrates dramatic forward-model speedups (ms vs. seconds) and robust inversion capabilities, including generalization to unseen structures and direct inversion mappings. The work highlights the potential of physics-ML to scale global seismic modeling and opens avenues for planetary seismic applications and broader earth science insights.

Abstract

Global seismic tomography, taking advantage of seismic waves from natural earthquakes, provides essential insights into the earth's internal dynamics. Advanced Full-waveform Inversion (FWI) techniques, whose aim is to meticulously interpret every detail in seismograms, confront formidable computational demands in forward modeling and adjoint simulations on a global scale. Recent advancements in Machine Learning (ML) offer a transformative potential for accelerating the computational efficiency of FWI and extending its applicability to larger scales. This work presents the first 3D global synthetic dataset tailored for seismic wavefield modeling and full-waveform tomography, referred to as the GlobalTomo dataset. This dataset is uniquely comprehensive, incorporating explicit wave physics and robust geophysical parameterization at realistic global scales, generated through state-of-the-art forward simulations optimized for 3D global wavefield calculations. Through extensive analysis and the establishment of ML baselines, we illustrate that ML approaches are particularly suitable for global FWI, overcoming its limitations with rapid forward modeling and flexible inversion strategies. This work represents a cross-disciplinary effort to enhance our understanding of the earth's interior through physics-ML modeling.

Figures (27)

• Figure 1: Overview of dataset.dataset is meticulously designed to tackle the pressing challenges associated with global seismic wavefield modeling and full-waveform inversion via cutting-edge physics-ml methodologies. In the forward modeling process, given specific source and velocity structures, the goal is to predict the wavefield at various time steps and the resulting seismograms at surface stations. The inversion process utilizes these seismograms as observational data to deduce the underlying velocity structures. Advanced ml techniques enhance these processes through neural operator learning and rapid automatic differentiation, substantially improving the efficiency of both forward modeling and inversion tasks.
• Figure 2: Qualitative forward modeling prediction by mlp in the Acoustic tier. A single slice is displayed. The source is located at $x = 0.00$ and $z = 0.80$. The background illustrates the velocity structure. The seismogram depicts the time series received by stations around the surface of this slice.
• Figure 3: Meta-analysis of forward modeling. (a) rl2, used to quantify error in wavefield prediction, shows increasing trends over time across baseline models. (b) Once trained, ml emulation significantly outpaces numerical simulations in speed, facilitating more iterations in inversion processes. (c) DeepONet, when trained on timesteps 1, 3, 5, and 7, struggles to generalize to intermediate timesteps such as 2, 4, 6, and 8. Incorporating physical constraints during training improves model performance on these denser timesteps.
• Figure 4: The inversion optimization process. Performance improves through optimization across 200 iterations. Higher degrees capture increasingly shorter-wavelength structures, enhancing model fidelity.
• Figure 5: Inversion with increasing starting points. Correlation between the inverted models and the ground truth strengthens as the number of starting points increases. Performance on different degrees shows a consistent trend.