Hyperparameter determination for GAN-based seismic interpolator with variable neighborhood search

Abstract

We propose an automatic global search algorithm based on the Variable Neighborhood Search (VNS) metaheuristic for tuning the hyperparameters of a generative adversarial network (GAN) seismic interpolator. We perform an exhaustive search to study the influence of each hyperparameter in the training process, and compare the proposed method with Random search and Bayesian Search. The seismic data set used for this study was synthetically modeled from a typical velocity model, estimated from a pre-salt field of the Brazilian cost. We also employ the proposed method with a real field data to show the importance and applicability of the search for optimum hyperparameters of GAN. The training data was constructed with decimated seismic data and the results were tested by comparing the reconstructed data with the original one. We performed two hyperparameter impact analyses: the first consists of an exhaustive grid exploration and the second consists of our proposed automatic exploration method using the VNS algorithm, comparing it with the other two algorithms. We concluded that the proposed method, which has a user-friendly usage, as it is almost parameter-free, can reach solutions with very good quality quickly, in any range of hyperparameter values. When compared with other methods of hyperparameter tuning, the one we propose proves to be better in the ease of configuration, while being efficient in the search process.