• neural network
• near-surface real data inversion
• geophysic inversions
• elastic full-waveform inversion
• big data analytics
• Rayleigh wave inversion
• residual neural network

This PhD work focuses on introducing big data analytics methodologies to the exploration geophysics field, with a particular interest in seismic inversions.

In this context, increasingly refined and accurate strategies have been successfully established in the wake of the technological advancements of recent decades. In particular, full-waveform inversion (FWI), which is an accurate inversion methodology for estimating subsurface properties, particularly the velocity field, has been successful as a reliable and robust method of geoseismic investigation. To address this problem, different optimization strategies have been adopted, including global optimization algorithms. This class of heuristic algorithms represents an excellent tool, guaranteeing good exploration capabilities at the expense of important computational time. Several strategies for mitigating these drawbacks have been proposed, and this work fits into this topic, proposing a hybrid strategy that combines a global optimization algorithm with an artificial intelligence algorithm.

The work starts from the preparatory experiments conducted to explore which technique can be more appropriate for developing the proposed methodologies. Specifically, various global optimization methods were analysed to identify the most suitable for FWI applications in this context. Six newly developed algorithms were compared, drawing inspiration from physical, biological or behavioural phenomena.

Furthermore, a study was carried out on the most effective parametrizations to be adopted for elastic forward modelling to guarantee the computational stability and limited dispersion errors of the calculated data without excessively increasing the computational times.

In the main section of this work, two different hybrid strategy implementations consisting of a combination of a genetic algorithm and a neural network (NN), used as the optimization engine of a global full-waveform inversion, are compared. In detail, to reduce the number of calculated models, the NN helps to predict some of the data misfit values necessary for the inversion procedure.
Three inversion problems are chosen on which the proposed methods are applied: a simple two-layer elastic 2D synthetic model, a very complex synthetic model constructed by interpolating two actual well logs, and finally real data acquired on a river levee in Colorno, Emilia Romagna, Italy.

The results show that the proposed methodology allows an effective reduction in computational times without affecting the effectiveness of the optimization process. This makes global optimizations, which are sometimes ignored due to the long computational times required, more attractive.

In this work, I limit the use of the NN as a sophisticated interpolator engine; however, this can easily be extended. In fact, the learning abilities of neural networks can facilitate inserting a priori conditions in the inversion, as well as favour using expensive and complex strategies such as joint inversions.

Finally, a minor project was developed in collaboration with the contractor ShearWater Geoservices regarding the elaboration of an effective designature strategy for appearance data.