GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 A data-driven supervised neural network approach for surface waves inversion: synthetc and feld data applicatons Felipe Rincón 1 , Sean Bert 1,2 , Mata Aleardi 1 , Eusebio Stucchi 1 1 University of Pisa, 2 University of Florence Introducton In near-surface applicatons, S-velocity models are commonly obtained through the analysis of the dispersion characteristc of the surface waves. One of the most popular approaches is the multchannel analysis of surface waves (MASW) which is a phase-velocity inversion method (Park et al., 1999). The forward operator involved in the computaton of the dispersion curves presents two strong assumptons: a 1D layered model and plane-waves. These limitatons strongly afect the capability of the method to account for lateral velocity variatons. To overcome these limitatons, it is imperatve to employ more sophistcated methods such as Full Waveform Inversion (FWI). FWI is an inverse problem that exploits the full informaton content of the seismic waveforms. Traditonally it is solved through deterministc approaches which seek to fnd a single best-ft model that explains the observed data. Even though this approach is computatonally efcient it heavily relies on a good startng model to reach convergence. The rapid advancements in algorithms and computng present an unprecedented opportunity for signifcant progress in seismic inversion, enabling the soluton of previously infeasible problems through data-driven approaches. A promising avenue of research involves establishing a direct inverse mapping from observed seismic waveforms to subsurface structures through the training of neural networks using paired data of seismic waveforms and corresponding velocity models (Wu et al., 2018). These approaches seek to leverage the power of deep learning to learn complex relatonships between seismic data and subsurface propertes, potentally revolutonizing the traditonal FWI methodology. However, the efcacy of learning-based methods stems from their ability to leverage vast amounts of high-quality training data, a challenge for seismic methods due to their high costs and confdentality concerns that limit the accessibility of seismic data. In this study we introduce a novel approach that combines a reparameterizaton of both the data and model parameters employing Discrete Cosine Transform (DCT) with neural networks to approximate the inverse operator. We tested our method in both synthetc and feld data from the InterPACIFIC project (Garofalo et al., 2016). Our objectve is to conduct tme-efectve training to generate S-velocity models from the data. The proposed model could serve as a startng point for a FWI frameworks, helping to mitgate the cycle skipping problem and reduce the number of iteratons to reach convergence.