Seismic refraction data analysis using machine learning and numerical modeling for characterization of dam construction sites

Abstract

Seismic refraction is a cost-effective tool to reveal subsurface compressional wave (P-wave) velocity. Inversion of traveltimes for estimating a realistic velocity model is a significant step in the processing of seismic refraction data. The results of the seismic data inversion are stochastic; thus, using prior information or complementary geophysical data can have a significant role in estimating the structural properties based on the observed data. Nevertheless, sufficient prior information or auxiliary data are not available in many geophysical sites. In such situations, developing advanced computational modeling is a vital step in providing primary information and improving the results. To this aim, a new inversion framework through hybrid committee artificial neural networks (CANNs) and the flower pollination (FP) optimization algorithm is introduced for inversion of refracted seismic traveltimes. Synthetic models generated by a forward-modeling approach are used to train the machine-learning model. Then, model parameters, such as the number of layers, thicknesses, and P-wave velocities, are predicted using a committee machine constructed based on several neural networks, which is achieved by averaging and stack generalization methods in which the latter method provides a better result. Then, the CANN results are used in the FP inversion algorithm to estimate the final model because it provides essential prior information on the number of layers and model parameters, which can be used in the FP searching algorithm. Our inversion procedure is tested on different synthetic data sets and applied at a dam site to determine the number of layers and their thicknesses. Our findings indicate a successful performance on synthetic and real data for automatic inversion of seismic refraction data.