Ground motion simulations rely heavily on subsurface velocity models that describe how seismic waves propagate through the Earth’s crust. These models range from simplified one-dimensional (1D) profiles to sophisticated three-dimensional (3D) descriptions that incorporate intricate geological features. Yet, despite advances in crustal modeling, the extent to which these diverse velocity models affect practical, engineering-relevant ground shaking predictions has remained inadequately quantified. The research led by scientists from the Southern University of Science and Technology addresses this knowledge gap by systematically assessing the predictive performance of nine representative velocity models, each constructed using different data sources, resolutions, and modeling methodologies.

Distinct from many previous studies that focus mainly on waveform matching or travel-time accuracy, the team shifts attention onto peak ground velocity (PGV)—a crucial metric for earthquake engineering and damage appraisal. PGV evaluates the maximum speed reached by ground shaking during an earthquake and has a direct correlation with potential structural damage and human perception of shaking intensity. By prioritizing PGV, the study bridges the gap between seismological modeling and practical applications in post-earthquake scenarios, providing insights directly relevant to emergency responders and structural engineers.

The methodology centers on the Luding earthquake, a significant seismic event that offers a realistic testing ground. Researchers employed a finite-fault rupture model derived from detailed seismic observations of the event. Using advanced numerical simulations conducted up to 1 Hz frequency, they calculated PGV distributions based on the nine distinct 3D crustal velocity models. Each model embodies unique assumptions about the Earth’s crustal properties, capturing differences in shallow velocity structure, topographic inclusion, and overall modeling strategies.

Results reveal that while most 3D velocity models can satisfactorily replicate observed PGV values at frequencies below 0.3 Hz, their accuracy diminishes at higher frequencies. This frequency-dependent performance underscores inherent challenges in capturing fine-scale geological heterogeneities and complex wave propagation phenomena. Notably, these 3D models consistently outperform traditional 1D models in predicting both the magnitude and spatial distribution of ground shaking, underscoring the necessity of incorporating three-dimensional structural complexity for realistic simulations.

However, the study identifies persistent discrepancies among individual 3D models. Some models exhibit systematic overestimations of shaking intensity, while others lean toward underestimation. These divergences are attributed primarily to variations in the representation of near-surface velocity structures, the degree to which surface topography is integrated, and differing conceptual approaches to model construction. Such differences highlight the sensitivity of simulation outcomes to nuanced geological characterizations and modeling choices.

A striking insight from the investigation is the demonstrated value of employing a multi-model ensemble approach. By averaging PGV predictions across all nine velocity models, the researchers attained a significantly reduced systematic bias and improved consistency with recorded ground motion observations. This multi-model averaging technique emerges as a pragmatic and robust strategy for rapid post-earthquake assessments, especially in settings where uncertainties in subsurface composition and structure may preclude reliance on any single velocity model.

The implications of these findings extend well beyond the immediate case study. They provide a quantitative framework for selecting and applying crustal velocity models in strong ground motion simulations, with direct relevance to seismic hazard analysis, earthquake engineering design, and emergency management protocols. Furthermore, the multi-model strategy proposed offers a pathway toward enhancing the reliability of shaking intensity forecasts in seismically active regions characterized by complex geology.

An additional contribution of this work lies in its emphasis on practical metrics like PGV rather than purely theoretical waveform concordance, aligning seismological research more closely with engineering and emergency response needs. By evaluating PGV at individual seismic stations as well as aggregating regional shaking patterns, the study presents a comprehensive suite of validation tests that ground motion modelers can adopt to benchmark their own velocity models and simulation frameworks.

In conclusion, this pioneering assessment of diverse 3D crustal velocity models in the context of the 2022 Luding earthquake offers critical insights into the interplay between subsurface modeling choices and ground shaking predictions. It reveals that while no single model is flawless, combining multiple models yields more reliable and actionable information. Such multi-model approaches are poised to play an increasingly central role in earthquake risk mitigation strategies, especially in tectonically complicated regions where seismic hazards are compounded by geological heterogeneity.

This study underscores the necessity of integrating cross-disciplinary expertise in seismology, geophysics, and computational modeling to confront the challenges of earthquake ground motion prediction. The findings advocate for continuous refinement of crustal velocity models, enhanced data acquisition for model construction, and adoption of ensemble modeling strategies as a best practice standard. Ultimately, these advances have the potential to save lives and minimize economic losses by informing more effective earthquake preparedness and response plans.

Future research prompted by this work may explore the integration of machine learning techniques for velocity model optimization, the coupling of strong motion simulations with urban infrastructure models, and real-time applications of multi-model approaches for immediate post-earthquake ground shaking estimation. As seismic hazard modeling progresses towards greater precision, studies like this forge a path toward more resilient societies capable of confronting the unpredictable power of earthquakes.

Subject of Research: Impact of Three-Dimensional Crustal Velocity Models on Strong Ground Motion Simulations

Article Title: Assessment of the Impact of Different 3D Crustal Velocity Models on Strong Ground Motion Simulations in the Sichuan-Yunnan Region

News Publication Date: Not explicitly provided

Web References: http://dx.doi.org/10.1007/s11430-025-1743-3

References:
Li T, Zhang W. 2026. Assessment of the impact of different 3D crustal velocity models on strong ground motion simulations in the Sichuan-Yunnan region. Science China Earth Sciences, 69(1): 348–365.

Image Credits: ©Science China Press

Keywords: strong ground motion, peak ground velocity, crustal velocity models, 3D seismic modeling, Luding earthquake, Sichuan-Yunnan region, seismic risk mitigation, earthquake engineering, numerical simulation, subsurface velocity structure, seismic intensity prediction

Earthquake early warning systems traditionally rely on a combination of seismic sensors and physical modeling of wave propagation to estimate the arrival times and intensities of damaging ground motions. However, these conventional methods often encounter latency issues and computational bottlenecks that limit their capacity to issue timely alerts, especially for large-magnitude events originating close to populated regions. The research team addresses these critical challenges by introducing a deep learning framework designed explicitly to forecast wavefield evolution rapidly and accurately using observational seismic data sequences.

At the core of their method lies a deep sequence-to-sequence neural network architecture, a class of models originally conceptualized for natural language processing tasks such as machine translation and text summarization. By ingeniously adapting this architecture to the spatiotemporal patterns inherent in seismic wavefields, the model learns to map historical ground motion sequences observed across sensor networks directly onto future wavefield states. This data-driven mapping bypasses the need for time-consuming numerical simulations based on physical equations, achieving unprecedented forecasting speed without sacrificing accuracy.

The researchers meticulously curated and preprocessed an extensive dataset of seismic wavefields generated from heterogeneous earthquake rupture scenarios simulated using high-fidelity physics-based models. This training corpus enabled the network to internalize the nonlinear dynamics governing wave propagation across complex geological structures, including phenomena such as wave scattering, diffraction, and mode conversions. As a result, the model develops an implicit physical intuition embedded within its latent space representations, allowing it to generalize to unseen seismic events with remarkable robustness.

One of the pivotal innovations in the team’s approach is the incorporation of temporal attention mechanisms within the sequence-to-sequence framework. This design allows the model to selectively weigh different time steps in the input sequence, effectively focusing on the most informative observations for predicting imminent ground motions. Such attention-driven forecasting not only improves accuracy but also enhances interpretability by highlighting critical seismic signal features that drive the evolution of the wavefield.

Benchmarking experiments demonstrated that this AI-powered wavefield forecasting system can predict imminent seismic wave arrivals at target locations several seconds faster than state-of-the-art physics-based simulators while offering comparable or superior predictive fidelity. This latency reduction is vital for early warning applications where every fraction of a second can translate into life-saving evacuation time or activation of automated safety protocols.

The study’s implications extend far beyond mere improvements in warning speed. By enabling rapid and precise spatial forecasting of ground shaking intensities, emergency managers and infrastructure operators could dynamically tailor response measures tailored to localized hazard footprints. For instance, automated control systems could initiate selective shutdowns of vulnerable facilities or adjust traffic signal timing along evacuation routes based on real-time predicted shaking distributions.

Furthermore, the data efficiency and scalability of the sequence-to-sequence model render it adaptable to various seismic networks worldwide, including those with sparse sensor coverage or noisy measurement environments. This flexibility addresses a long-standing limitation in earthquake early warning deployment across regions with limited instrumentation or challenging geophysical conditions.

Despite these promising outcomes, the authors acknowledge ongoing challenges requiring further investigation. Integrating the AI-based forecasts within operational early warning frameworks necessitates rigorous validation under real earthquake scenarios and seamless interoperability with existing alert dissemination mechanisms. Additionally, understanding the model’s failure modes and ensuring robustness against rare or extreme seismic phenomena remain crucial for trustworthy system performance.

The study also opens exciting interdisciplinary research directions at the confluence of seismology, machine learning, and risk management. Future work could explore hybrid models that combine physics-based constraints with deep learning to enforce physically consistent and interpretable forecasts. Incorporating real-time data assimilation techniques may further refine prediction accuracy as events unfold, enabling truly dynamic and adaptive early warning capabilities.

By harnessing the power of deep sequence-to-sequence learning, this novel wavefield forecasting paradigm marks a significant leap toward more proactive and predictive earthquake resilience. As urban populations grow and infrastructure becomes increasingly interconnected, the societal value of such technological breakthroughs cannot be overstated. Rapid, reliable earthquake early warning promises to transform how communities anticipate, prepare for, and respond to one of nature’s most formidable hazards.

In conclusion, the innovative methodology proposed by Lyu et al. exemplifies the transformative potential of AI-driven waveform forecasting to enhance early warning systems globally. Their work eloquently demonstrates how data-centric, physics-informed machine learning architectures can complement and transcend traditional seismic modeling approaches, delivering faster, actionable insights in the critical seconds preceding destructive ground shaking. As this technology matures and integrates with broader disaster management frameworks, it holds the promise to save lives, reduce economic losses, and fundamentally elevate earthquake hazard preparedness across diverse seismic regions.

The research community eagerly anticipates further empirical validations, optimization, and operational deployment of such AI-powered forecasting tools. This cutting-edge fusion of seismology and deep learning heralds a new era in earthquake science—one where rapid wavefield prediction via advanced neural networks becomes a cornerstone of next-generation hazard mitigation and societal resilience strategies worldwide.

Subject of Research: Earthquake early warning and seismic wavefield forecasting using deep sequence-to-sequence learning models.

Article Title: Rapid wavefield forecasting for earthquake early warning via deep sequence to sequence learning.

Article References:
Lyu, D., Nakata, R., Ren, P. et al. Rapid wavefield forecasting for earthquake early warning via deep sequence to sequence learning. Nat Commun (2025). https://doi.org/10.1038/s41467-025-65435-2

Image Credits: AI Generated