On the implementation of large-scale integral operators with modern HPC solutions -- Application to 3D Marchenko imaging by least-squares inversion

Abstract: Numerical integral operators of convolution type form the basis of most wave-equation-based methods for processing and imaging of seismic data. As several of these methods require the solution of an inverse problem, multiple forward and adjoint passes of the modelling operator must be performed to converge to a satisfactory solution. This work highlights the challenges that arise when implementing such operators on 3D seismic datasets and it provides insights into their usage for solving large systems of integral equations. A Python framework is presented that leverages libraries for distributed storage and computing, and provides an high-level symbolic representation of linear operators. To validate its effectiveness, the forward and adjoint implementations of a multi-dimensional convolution operator are evaluated with respect to increasing size of the kernel and number of computational resources. Our computational framework is further shown to be suitable for both classic on-premise High-Performance Computing and cloud computing architectures. An example of target-oriented imaging of a 3D synthetic dataset which comprises of two subsequent steps of seismic redatuming is finally presented. In both cases, the redatumed fields are estimated by means of least-squares inversion using the full dataset as well as spatially decimated versions of the dataset as a way to investigate the robustness of both inverse problems to spatial aliasing in the input dataset. We observe that less strict sampling requirements apply in three dimensions for these algorithms compared to their two dimensions counterparts. Whilst aliasing introduces noise in the redatumed fields, they are however deprived of the well-known spurious artefacts arising from incorrect handling of the overburden propagation in cheaper, adjoint-based redatuming techniques.