Inverse theory, artificial neural networks

Abstract

Geophysical inversion is an essential tool to help geophysicists solve practical environmental, engineering, or exploration problems. This is particularly relevant for situations that require detailed quantitative subsurface information and data. Traditional approaches to inversion are many and varied, and include model perturbation/comparison, linearization usually of a nonlinear geophysical model, together with a variety of stochastic methods including Monte Carlo, simulated annealing, and genetic algorithms. These approaches do not perform well in the presence of noise, are very inefficient and computing intensive, require subjective tuning parameters and a priori knowledge, and/ or produce results that are nonunique. Alternative geophysical inversion methods, based on artificial neural networks (ANNs), have been investigated over the past 15–20 years, to try and overcome the disadvantages of the traditional approaches. These ANN inversion methods have now found considerable success across many different geophysical domains, including seismic attributes, velocity fields, and tomography, to potential field inversion, archaeology, and remote sensing. With the continuous introduction of improved ANN paradigms and training algorithms, coupled with the ever-increasing power of digital computing hardware (including Field Programmable Gate Arrays [FPGAs]), there is no doubt that ANNs have a very positive future in the exciting and challenging world of geophysical inversion.