We perform synthetic tests of a new genetic algorithm (GA) for 2-D velocity estimation from refraction time data. Based on evolutionary statistics, GAs are Monte-Carlo optimizations designed for nonlinear, multi-modal, and large-scale problems. Each individual in our GA is a trial 120-by-46 element 2-D velocity grid. An initial population of 200 models undergoes binary tournament selection, 2-D block crossover, and seismologically-constrained mutation to generate a new population of 200, selected through forward travel-time modeling to better fit the data. We iterate this process until convergence. Like all Monte Carlo methods, GAs are computationally-intensive, but are adaptable to parallel implementation on networked workstations.
We tested our GA on two laterally-variable synthetic models, using the geometry of a real refraction survey; one with a layered structure plus a high-velocity intrusion, and the other substituting a low-velocity magma chamber. We add random errors to the data in proportion to the pick qualities of the real data. We find that a GA run of 100 generations can find a suite of acceptable models similar to the original model. We use the suite's standard deviation at each grid point to assess resolution. Our results show the velocity oscillations typical of inversions of data sets having anisotropic ray coverage. We plan to include S-wave refraction picks and surface wave phase velocities in our data set, which should constrain the oscillations and allow estimation of Poisson's ratio structure.