Answers are grounded in this paper and cite their sources.
Summary Deep-learning-based seismic velocity inversion is often affected by observational noise and insufficient data–model consistency. To address these issues, we propose a forward–inverse dual-constrained deep learning framework for shear-wave velocity (Vs) inversion from surface-wave dispersion data. The framework consists of a convolution-free inversion subnetwork based on multi-head self-attention and a learned forward-modelling subnetwork. The learned forward-modelling subnetwork links the predicted Vs model back to the dispersion-data domain, allowing the inversion subnetwork to be constrained simultaneously by a velocity-model consistency term and a data-domain consistency term. Synthetic tests show that, compared with a single-constraint network using the same backbone, the dual-constrained network yields lower model-domain errors. After recalculating dispersion with disba, the phase-velocity error decreases, the group-velocity error is slightly higher, and the combined phase–group RMSE decreases only slightly. When applied to 11,397 Rayleigh-wave dispersion curves across East Asia, the trained network yields a regional Vs structure that is broadly consistent with the major tectonic features and shows good agreement in its principal structural patterns with the regional Vs model of Li et al. (2022). In the synthetic test, the dual-constrained framework addresses both Vs-model recovery and joint phase- and group-velocity dispersion fitting, while retaining the ability to process regional dispersion curves rapidly after training.
No takes yet. Share an insight, caveat, or question.
李建平 et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: