Global optmization methods to solve DC-resistivity inverse problems with flexibility in constraint incorporation
Global Optimization, Simulated Annealing, Genetic Algorithms, Particle Swarm Optimization, Resistivity
Inversion in DC-resistivity is an ill-posed inverse problem because different realizations of the same model might satisfy approximately the same data fitting criterium. It is therefore necessary to use constraints to obtain unique and / or stable solutions to small perturbations in the measurements. However, in general, the introduction of constraints has been restricted to cases of differentiable constraints, which can be treated with local optimization algorithms. 1D and 2D modeling in DC-resistivity is computationally inexpensive, allowing the use of global optimization methods (GOMs) to solve 1.5D and 2D inverse problems with flexibility in constraint incorporation. Changes in the cost function, either in the constraints or data fitting criteria, can be easily performed, since each term of the cost function is properly normalized to allow the approximate invariance of the Lagrange multipliers. GOMs have the potential to support a computational environment suitable for quantitative interpretation in which the comparison of solutions incorporating different constraints is one way of inferring characteristics of the actual distribution of the underground resistivity. In this work, we developed: (i) comparison of the performances of the Simulated Annealing (SA), Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) methods to solve the 1.5D inverse problem in DC resistivity using synthetic and field data; (ii) an inversion approach based on particle swarm optimization (PSO) to solve the 2D DC-resistivity inverse problem; (iii) exploration of several constraints in the variation of log-resistivity, including spatial continuity in both L1 and L2 norms, total variation and sparsity constraints using discrete cosine and Daubechies bases. In addition, we explore the minimum inertia constraint, including the case of using the Earth’s surface as the target axis, to impose the concentration of resistive or conductive materials along target axes. The main results of the comparison for the 1.5D case are: a) all methods reproduce quite well the resistivity distribution of synthetic models, b) PSO and GA are very robust to changes in the cost function and SA is comparatively much more sensitive, c) PSO first and GA second present the best computational performances, requiring smaller number of forwarding modeling than SA, and d) GA shows the best performance with respect to the final attained value of the cost function and its standard deviation, whilst SA has the worst performance in this aspect. Equally important for both 1.5 and 2D cases, from the stopping criteria of the PSO algorithm results not only the best solution but also a cluster of suboptimal quasi-solutions from which uncertainty analyses can be performed. As a result, the interpreter has freedom to perform a quantitative interpretation process based on a feedback trial-and-error inversion approach, in a similar manner he/she has when using a friendly forward modeling software, being capable of driving the solution to incorporate his/her conceptions about the geologic environment, besides appraising data fitting and stability of the obtained solutions. We present both synthetic and field data examples for all inversion cases.