Velocity inversion in cross-hole seismic tomography by counter-propagationneural network, genetic algorithm and evolutionary programming techniques

The disadvantages of conventional seismic tomographic ray tracing and inver sion by calculus-based techniques include the assumption of a single ray pa th for each source-receiver pair, the non-inclusion of head waves, long com putation times, and the difficulty in finding ray paths in a complicated ve locity distribution. ii ray-tracing algorithm is therefore developed using the reciprocity principle and dynamic programming approach. This robust for ward calculation routine is subsequently used for the cross-hole seismic ve locity inversion. Seismic transmission tomography can be considered to be a function approxim ation problem; that is, of mapping the traveltime vector to the velocity ve ctor. This falls under the purview of pattern classification problems, so w e propose a forward-only counter-propagation neural network (CPNN technique for the tomographic imaging of the subsurface. The limitation of neural ne tworks, however, lies in the requirement of exhaustive training for its use in routine interpretation. Since finding the optimal solution, sometimes from poor initial models, is the ultimate goal, global optimization and search techniques such as simula ted evolution are also implemented in the cross-well traveltime tomography. Genetic algorithms (GA), evolution strategies and evolutionary programming (EP) are the main avenues of research in simulated evolution. Part of this investigation therefore deals with GA and EP schemes for tomographic appli cations. In the present work on simulated evolution, a new genetic operator called 'region-growing mutation' is introduced to speed up the search proc ess. The potential of the forward-only CPNN, GA and EP methods is demonstrated i n three synthetic examples. Velocity tomograms of the first model present p lausible images of a diagonally orientated velocity contrast bounding two c onstant-velocity areas by both the CPNN and GA schemes, but the EP scheme c ould not image the model completely. In the second case, while GA and EP sc hemes generated an accurate velocity distribution of a faulted layer in a h omogeneous background, the CPNN scheme overestimated the vertical displacem ent of the fault. One can easily identify five voids in a coal seam from th e GA-constructed tomogram of the third synthetic model, but the CPNN and EP schemes could not replicate the model. The performances of these methods a re subsequently tested in a real field setting at Dhandadih Colliery, Ranig anj Coalfields, West Bengal, India. First arrival traveltime inversion by t hese algorithms from 225 seismic traces revealed a P-wave velocity distribu tion from 1.0 to 2.5 km s(-1). A low-velocity zone (1.0 km s(-1)), the posi tion of a suspected gallery in the Jambad Top coal seam, could be successfu lly delineated by CPNN, GA and EP schemes.