Inversion of potential field data by genetic algorithms

We present a genetic algorithm that simultaneously generates a large number of different solutions to various potential field inverse problems. It is shown that in simple cases a satisfactory description of the ambiguity domain inherent in potential field problems can be efficiently obtained by a simple analysis of the ensemble of solutions. From this analysis we can also obtain information about the expected bounds on the unknown parameters as well as a measure of the reliability of the final solution that cannot be recovered with local optimization methods. We discuss how the algorithm can be modified to address large dimensional problems. This can be achieved by the use of a 'pseudo-subspace method', whereby problems of high dimensionality can be globally optimized by progressively increasing the complexity and dimensionality of the problem as well as by subdividing the overall calculation domain into a number of small subdomains. The effectiveness and flexibility of the method is shown on a range of different potential field inverse problems, both in 2D and 3D, on synthetic and field data.