Multi-Level Transfer Learning for Optimal Features Prediction
Muhammad Sajid · International Petroleum Technology Conference · 2024
Abstract The quality of the final predictions in machine learning (ML) depends on a multitude of factors, and among these factors, the generalization of features within the input dataset plays a pivotal role. This generalization is particularly crucial when dealing with seismic data, as the complexity of the underlying geological structures and the seismic characteristics themselves exhibit substantial variability across different geological study fields. Such variations pose unique challenges when it comes to model training and achieving optimal predictions for specific seismic features. Considering these challenges, the research proposes a comprehensive approach that involves the application of three stages of transfer learning. This approach has been designed to facilitate the seamless convergence of the model towards an optimal level of adaptability for the extraction of seismic features. Each step of transfer learning in this method incrementally enhances the model's ability to accommodate the idiosyncrasies of seismic data from diverse study fields, ultimately leading to the desired seismic feature prediction. The research seeks to improve the overall accuracy and robustness of seismic feature prediction across varied geological contexts, thereby contributing to the advancement of seismic data analysis and geological interpretation.