Development of an ANN-Based Technique for Inversion of Seismic Refraction Travel Times

Abstract

Non-uniqueness in the inversion of seismic data can be considered the main challenge for the application of such data. Prior information, such as downhole data, can be used to control this problem. However, in most cases, prior information is not available; accordingly, geophysicists/analysts have to suppose a primary model for the observed data and then find the final adequate layered earth model through trial and error. In this study, a new technique was developed based on the artificial neural network (ANN) for the inversion of seismic refraction data in the absence of prior information. In this regard, a sequential multilayer perceptron (SMLP) was proposed, which integrates the sequential information of the model parameters to predict a reasonable layered earth model. In fact, at first, a multilayer perceptron (MLP) (First-MLP) was trained by synthetic data; then, a layered earth model, i.e., the primary model, was predicted for the observed data. Next, using the primary model, a range for each of the model parameters, i.e., thickness and P-wave velocity, for each layer was defined. Subsequently, new synthetic samples were generated based on the determined ranges. Finally, using another MLP (Second-MLP), which was trained by the new synthetic samples, the final model for the observed data was estimated. The proposed method was also tested by employing different synthetic data with and without noise. Moreover, the SMLP inversion technique was used to analyze the experimental seismic refraction dataset at a dam construction site. The results for both synthetic and experimental data confirmed the reliability of the proposed SMLP inversion technique.