TL;DR
DispFormer is a novel transformer-based architecture that solves a long-standing challenge in geophysics: inverting surface wave dispersion curves into subsurface shear-wave velocity ($v_s$) profiles using variable-length, noisy, or incomplete data. By processing each period independently and incorporating a physics-informed dynamic depth alignment strategy, it delivers highly accurate subsurface models in both zero-shot and few-shot settings, outperforming traditional global search methods while cutting inference time down to seconds.
What Problem Does This Paper Address?
Mapping the Earth's subsurface structure is vital for earthquake hazard assessment, resource exploration, and geological studies. Geophysicists often rely on surface wave tomography, specifically inverting dispersion curves (which describe how seismic wave velocity changes with frequency/period) to estimate 1-D shear-wave velocity ($v_s$) profiles.
However, traditional inversion techniques suffer from severe bottlenecks: linearized methods are highly sensitive to the initial "guess" model, while global search algorithms (like Particle Swarm Optimization) are computationally crushing.
While deep learning offers a faster alternative, standard neural networks (like CNNs or ResNets) require fixed-length inputs. Real-world seismic data is rarely clean or uniform; it frequently features varying period ranges, missing data points, and low signal-to-noise ratios.
Methods
To break free from the fixed-length constraint, the authors introduced DispFormer, a transformer-based neural network.
Instead of treating the entire dispersion curve as a single fixed image or vector, DispFormer processes the data point at each period independently. Each period's data (period, phase velocity, and group velocity) is encoded via shared-weight dense layers and combined with position embeddings to preserve temporal/frequency relationships. Transformer blocks then capture the complex inter-period correlations required to model depth-dependent velocity structures.
Crucially, the authors integrated a dynamic depth alignment strategy to enforce physical consistency. Using empirical wave physics formulas, the network dynamically calculates the minimum and maximum sensitive depths based on the wavelengths of the observed dispersion data. This effectively narrows the mathematical solution space to physically plausible features.
DispFormer uses a pre-training and fine-tuning workflow. It is first trained on a massive global synthetic dataset and can then be deployed directly (zero-shot) or fine-tuned with minimal regional data (few-shot).
Datasets
The model was trained and validated across synthetic and real-world datasets:
- Global Pre-training Dataset: A completely synthetic dataset consisting of 40,962 1-D $v_s$ profiles extracted from the LITHO1.0 global crust and lithosphere model, spanning depths up to 200 km.
- Regional Synthetic Datasets (Validation): The Central and Western US Dataset (CWD; 6,803 profiles) and the Continental China Dataset (CCD; 4,527 profiles).
- Real-World Field Dataset: 12,705 observed dispersion curves extracted from the China Seismological Reference Model (CSRM).
Key Insights and Ideas
- The Power of Period-Independence: Treating dispersion data points like words in a sentence allows the transformer architecture to naturally handle missing data gaps and arbitrary curve lengths without needing destructive preprocessing like interpolation or padding.
- Zero-Shot as an Initial Model Generator: Even without any regional training data, the zero-shot DispFormer generates initial subsurface models that significantly outperform standard baseline interpolation methods, making it an excellent "plug-and-play" front-end for traditional inversion workflows.
- Physics-Informed Constraints Work: Incorporating wavelength-to-depth sensitivity directly into the loss function as a soft constraint drastically reduced Mean Absolute Error (MAE) and improved the model's geological explainability.
Key Results
- Zero-Shot Accuracy: On the synthetic CWD and CCD datasets, zero-shot DispFormer cut the MAE down to ~73 m/s, compared to ~234 m/s for the standard interpolated LITHO1.0 reference model.
- Few-Shot Superiority: When fine-tuned with a tiny fraction of regional data (just 10 samples for CWD or 36 for CCD), DispFormer achieved results comparable to traditional Particle Swarm Optimization (PSO). When provided with slightly more data (108–180 samples), it decisively outperformed PSO.
- Massive Computational Speedup: Traditional PSO global search required roughly 159 hours to process the CWD dataset. In contrast, fine-tuning DispFormer took less than 4 minutes, and its inference time for the entire dataset was approximately 1 second.
- Extreme Robustness: The model maintained stable performance even when subjected to simulated real-world degradation, showing strong resilience against up to 10% Gaussian white noise and up to 50% missing data points.
Limitations and Future Work
While DispFormer marks a major step forward, the authors noted a few limitations:
- Resolution Constraints: The model currently utilizes a uniform 0.5 km layer thickness, which limits its ability to resolve highly fine-scale structures essential for shallow engineering or near-surface urban exploration.
- Geological Complexity: The global pre-training data (LITHO1.0) does not fully capture extremely sharp or complex velocity contrasts found in unique shallow crust environments.
Future Directions: The authors plan to reconstruct the pre-training datasets with finer, variable layer thicknesses and extend DispFormer to support multi-physics joint inversions—such as integrating surface wave data with receiver functions, gravity, or magnetotelluric data.
Summary updated