GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 PROBABILISTIC APPROACH TO FULL-WAVEFORM INVERSION OF SURFACE WAVES: A REAL DATA APPLICATION S. Bert 1,2 , M. Aleardi 1 , E. Stucchi 1 1 Department of Earth Sciences (University of Pisa, Italy) 2 Department of Earth Sciences (University of Florence, Italy) Introducton Surface waves play a crucial role in near-surface geophysics, ofering a non-invasive way to determine the elastc propertes of near-surface sediments: this turns out to be of fundamental importance, for example, for geotechnical site characterizaton. This analysis started with the spectral analysis of surface waves (SASW) and became increasingly popular afer the introducton of the multchannel analysis of surface waves (MASW). The main limitaton of these approaches is the reliance on a 1D layered model assumpton, making them less efectve in the presence of substantal lateral heterogeneity or when dealing with multmodal dispersion paterns in the context of low-velocity layers and strong velocity contrasts. The advent of increased computatonal power in recent decades has made it possible the applicaton of the full-waveform inversion (FWI) approach, which exploits the full informaton content of the recorded seismogram to infer high-resoluton estmatons of subsurface acoustc or elastc parameters. While acoustc FWI is commonly employed for imaging complex subsurface structures, it falls short in near-surface seismic studies due to the prevalence of surface waves. In this work, our focus shifs to multparameter elastc FWI, aiming to construct P-wave and S-wave velocity models for near surface sediments. The inclusion of surface waves in the wavefelds increases the nonlinearity of FWI, elevatng the risk of the local approach getng stuck in some local minima of the usual L2 norm error functon. In this context, the inversion outcomes become strongly dependent on the startng model. Although global optmizaton methods can mitgate this issue, they come at the cost of signifcantly increased computatonal demands (Lamuraglia et al. 2022). Trying to overcome these issues, we propose a Bayesian inference framework for elastc FWI. Diferently from the local approach, the proposed method provides a comprehensive evaluaton of the uncertainty afectng the retrieved soluton through the so-called posterior probability density functon (PPD) in the model space. Based on the Bayes theorem, the PPD incorporates the informaton coming from both the prior knowledge on the model parameters and the recorded seismic data but, in case of nonlinear forward modelling, a sampling technique needs to be adopted to approximate this density functon. In our case a Markov Chain Monte Carlo (MCMC) sampling strategy is used to numerically evaluate the statstcal propertes of the PPD. However, challenges arise in the form of the convergence rate dependency on the proposal distributon and the diminished sampling ability in high dimensional spaces, known as the curse of dimensionality. To tackle these issues, we introduce a gradient-based Markov Chain Monte Carlo (GB-MCMC) method where the proposal distributon is constructed by the local gradient and the Hessian of the negatve log posterior, and we also reduce the dimensionality of the problems making use of the