AI Meets Physics in Surface-Wave Seismology: Faster Workflows, Trusted Results
Credit: Big Data and Earth System
Surface-wave methods link frequency to depth, which makes them ideal for imaging near-surface structure. The catch: traditional workflows depend on manual picking and iterative inversion, so they're slow and hard to scale. AI changes the speed curve, but speed without physical consistency can mislead. A recent review in Big Data and Earth System (DOI: 10.1016/j.bdes.2025.100039) tackles exactly this tension.
What the review covers
Researchers from Zhejiang University of Technology, Zhejiang University, and Anhui University of Science and Technology survey AI across the surface-wave workflow: automated dispersion extraction, deep-learning-based inversion, physics-guided modeling, and explainable AI. They compare data-driven sensitivity patterns with classical kernels to test whether models are learning real physics or just correlations. The result is a clear map of strengths, gaps, and practical fixes.
Why this matters for your lab
Manual dispersion picking and iterative inversion don't scale to dense arrays or near-real-time monitoring. Neural networks can extract dispersion curves automatically and invert them into shear-wave velocity models in seconds once trained. That makes high-density surveys and frequent updates realistic, even for distributed acoustic sensing (DAS) lines. For background on DAS, see the USGS overview here.
The reliability gap
Speed isn't the only metric. The review shows that some networks produce Jacobians that don't align with classical sensitivity kernels. When depth-frequency relationships are off, models lean on dataset quirks instead of physics. That's risky in poorly constrained depth ranges and heterogeneous settings.
What actually works
- Automated dispersion analysis: Deep models reliably track dispersion ridges in noisy, multimode data, reducing human bias.
- Fast inversion: Once trained, networks generate Vs profiles much faster than traditional optimization, enabling large-scale imaging.
- Physics-guided designs: Injecting constraints-geological priors, monotonicity, smoothness, or PDE-based regularization-improves stability and interpretability.
- Explainable AI: Comparing network sensitivities with physical kernels exposes where a model is trustworthy-and where it isn't.
Case note: detecting karst cavities
The review includes a case where AI-assisted feature analysis flags karst cavities from velocity models more consistently than manual inspection. The takeaway: AI can sharpen detection of subtle anomalies, but confidence should track where physical sensitivities support the prediction.
Physics-guided AI in practice
Physics-informed or physics-guided approaches keep models grounded while preserving speed. They can enforce dispersion physics during training or constrain outputs to realistic subsurface profiles. This hybrid path offers a practical balance: automation that respects wave propagation.
Applications you can scale now
- Urban hazard assessment and infrastructure planning
- Groundwater monitoring and environmental surveys
- Dense arrays and DAS networks for near-real-time updates
Interpretable models help teams quantify uncertainty instead of glossing over it. That reduces overconfidence and improves decisions under time pressure.
Implementation checklist
- Benchmark dispersion pickers on noisy, multimode datasets with known truth.
- Compare network-derived Jacobians with classical kernels before deployment.
- Add physics: monotonic depth trends where appropriate, bounds on Vs, and kernel-informed loss terms.
- Use uncertainty estimates (ensembles, MC dropout, or quantile outputs) to flag weakly constrained depths.
- Validate on geologies outside your training set; track failure modes, not just mean error.
- Standardize datasets and reporting so results are comparable across teams and sites.
Standards and next steps
The review argues for shared datasets, open benchmarks, and architectures that encode physics by design. As these mature, AI-driven surface-wave imaging can move from lab demos to routine, reliable workflows in Earth science and engineering.
Further reading
- Distributed Acoustic Sensing overview (USGS): link
- Physics-informed neural networks (introductory review): arXiv:1711.10561
Reference
Big Data and Earth System, November 28, 2025. DOI: 10.1016/j.bdes.2025.100039
Funding
National Natural Science Foundation of China (Grant No. 42304155) and Zhejiang Provincial Natural Science Foundation of China (Grant No. LMS25D040001).
Want to upskill your team on AI methods?
Explore curated AI coursework for technical professionals: Latest AI courses.
Your membership also unlocks:
