Block sparse Bayesian learning-based prestack seismic inversion with the correlation of velocities and density

Abstract

Prestack seismic inversion has been widely used around the seismic exploration. It can precisely output the elastic properties of layers in subsurface, e.g., P-wave velocity ($V_{p}$), S-wave velocity ($V_{s}$), and density ($\rho$). These are utilized further to extract many reservoir properties, like saturation and porosity, which are very helpful for the successful oil field development. The accuracy of prestack inversion result could play a critical role for the evaluation of reservoir characterization and the quantitative interpretation. It has been observed the existing relationships among velocities and density ($V_{p}$, $V_{s}$, and $\rho$) lead to the correlations of three reflectivities ($R_{p}$, $R_{s}$, and $R_{\rho }$). In this paper, we establish a new formulation for amplitude-versus-angle (AVA) inversion incorporating these correlations. A machine learning technique—block sparse Bayesian learning, has been implemented as the inversion engine to solve $R_p$, $R_s$, and $R_\rho$ and to estimate the correlations described as the covariance matrix automatically. Due to the contribution of relationship between velocities and density, the performance of proposed AVA inversion is superior to the conventional technique in denoising and highlighting small-scale reflections. Three reflectivities are finally converted into velocities and density via an optimal multi-trace algorithm L-BFGS, which can mighty mitigate the lateral discontinuity of inverted results. The proposed approach has been tested on the synthetic examples. It shows a good consistence between inverted and true elastic properties. Field data test with the seismic profiles in Ordos basin area has demonstrated its high feasibility and efficiency for practical applications.