Full Waveform Inversion of common offset GPR data using a fast deep learning based forward solver

Abstract

Electromagnetic (EM) forward solvers, such as the finite-difference time-domain (FDTD) method are an essential part for the interpretation of the GPR data. Their drawback is that they are still computationally expensive algorithms and not easily applicable for simulating real scenarios in the absence of high performance computing (HPC). Machine learning (ML) can provide a solution to this problem for specific applications by providing near real time solutions to the forward problem. In this paper, we have developed an ML-based forward solver that is used in full-waveform inversion (FWI) schemes and is applied to concrete slab scenarios. A model of a real GPR transducer was used in the simulations and as a result the algorithm can be used for the inversion of real data. The coupled ML solver/FWI algorithm was tested with both synthetic and real data to assess its performance. Although the algorithm was tuned for a concrete slab case, it can be adjusted and applied to different GPR applications.