A data-driven transdimensional approach to include lateral constraints on 2D target-oriented AVA inversion

Seismic inversion aims to infer subsurface properties from processed seismic data; since these are often ill-conditioned procedures, numerous strategies can be investigated. To date, currently adopted procedures assume an a priori structural knowledge of the investigated area and impose such constraints on the recovered solution. To overcome this shortcoming, we apply a transdimensional reversible jump-Markov chain Monte Carlo (Rj-McMC) algorithm to solve the interval-oriented amplitude versus angle (AVA) inversion on 2D synthetic seismic data. This approach considers the model parameterisation as an unknown, together with the elastic properties of the investigated area. The algorithm samples models discretised in Voronoi cells characterised by similar AVA responses. The elastic values associated with each Voronoi cell are obtained taking the average of the elastic properties of the Common Dip Points (CDPs) falling within it. This data-driven approach, therefore, needs no external assumption over the investigated area and ensures an automatically inferred strategy to include lateral variability of data inside the inversion kernel. We compare results obtained with a standard Bayesian approach for different signal-to-noise ratios (SNR), showing how the increase of random noise contaminating the data strongly affects the linear approach, while the Rj-McMC generates model predictions in accordance with the true model, producing more reliable results.