Machine and Deep Learning Seminar

Abstract:

Seismic processing commonly deals with eliminating collected seismic data's inherent multiples. Traditional methods like Radon transform-based techniques are time-consuming and complex. We introduce a deep learning-based alternative in this work that simplifies the process while maintaining competitive performance. Our model, trained only on synthetic data, works well on complex field data. We ensure it preserves data characteristics without over-smoothing while removing multiples. We also conduct a thorough analysis, focusing on hyperparameter effects and Bayesian performance assessment, to provide insights into the model's inner workings.