Neural Sensitivity and Wave Tangent Kernels Explain NeurFWI Convergence
A new study from arXiv (2605.14370) introduces the neural sensitivity kernel (NSK) and wave tangent kernel (WTK) to theoretically analyze neural reparameterized full-waveform inversion (NeurFWI). NeurFWI uses neural networks to represent wave equation parameters, reducing dependence on high-quality initial models and wavefield data but suffering from slow high-resolution convergence. The NSK and WTK frameworks show that the neural tangent kernel (NTK) adaptively modulates original sensitivity and wave tangent kernels, producing spectral filtering, gradient wavenumber modulation, and wave frequency bias. These effects connect convergence behavior across model and data domains, clarifying the underlying mechanism of NeurFWI.
Key facts
- Study published on arXiv with ID 2605.14370
- Introduces neural sensitivity kernel (NSK) and wave tangent kernel (WTK)
- NeurFWI reduces reliance on high-quality initial models and wavefield data
- NeurFWI has slow high-resolution convergence
- NTK adaptively modulates original sensitivity and wave tangent kernels
- Modulation leads to spectral filtering effect
- Modulation leads to gradient wavenumber modulation
- Modulation leads to wave frequency bias
Entities
Institutions
- arXiv