Semi-supervised multi-task learning for seismic interpretation

Abstract

Semi-supervised multi-task learning for seismic interpretation. Saleem Asghar Department of Resources and Environmental Engineering Graduate school, Hanyang University Traditional seismic data interpretation methods in oil field exploration, relying on manual interpretation and rule-based algorithms for subsurface geological structure analysis, face inherent limitations. Specifically, traditional horizon tracking and lithofacies classification methods are subjective, iterative, and time-consuming. These approaches introduce bias, inconsistency, and potential inaccuracies. Horizon tracking, for instance, is hindered by its subjective and time-consuming nature, while lithofacies classification struggles with overlapping facies and high data dimensionality. Moreover, the limited availability of labeled training data constrains the capacity of machine learning (ML) models employed in performing these tasks to effectively address the associated challenges. ML models, adept at discerning intricate patterns from data, offer a promising solution to the limitations of traditional methods. However, their success relies heavily on the availability of extensive training datasets. In geophysical problems, obtaining labeled training data is a challenging and costly endeavor, particularly in subsurface geological studies. This underscores the need for novel methodologies capable of providing a more scalable and cost- effective approach. This thesis explores multi-task learning (MTL) methodologies to address the challenges paucity of training data in horizon tracking and lithofacies classification. By leveraging MTL’s capacity to jointly learn correlated tasks and optimize limited labeled data, this research addresses the lack of training data and the class imbalance problem by utilization of MTL in a semi-supervised setting different from the conventional data augmentation methods. The first study “multi-task learning for automatic horizon tracking with limited labeled data” demonstrates that a MTL based ML model trained on a small amount of labeled data can achieve significant enhancement in prediction accuracy and generalization by leveraging unlabeled data through joint learning. The MTL framework, tailored for horizon tracking, shows its effectiveness in handling correlated tasks. Through shared representations between multiple tasks, the model achieves enhanced performance with limited training data demonstrating its adaptability to geological complexities, verified in the North Sea F3 oil field. The second study “multi-task learning for facies classification (MTLFC)” proposes innovative methods to accurately classify facies by harnessing the rich resource of unlabeled seismic data and sparse labeled data. The proposed MTLFC employs representation learning to benefits from unlabeled data. The unlabeled data also acts as a constraint and regularization method in training phase. Moreover, by integrating a custom loss function based on label smoothing, the proposed MTL model assigns relatively less confidence to the majority classes and minimizes the confidence of the model in excessive classes. Employing the proposed method on data from the Vincent oil field yielded more accurate facies classification results compared to traditional methods, conventional machine learning approaches, and supervised learning methods. Based on the extensive experiments and results of horizon tracking and facies classification, I conclude that the proposed methods represent advanced tools for practitioners to interpret seismic data efficiently and accurately, even with limited labeled data. This is achieved through the synergistic combination of relevant tasks and workflows. Moreover, the applicability of the proposed approach extends beyond horizon tracking and facies classification, encompassing diverse potential applications such as inversion, denoising, interpolation, and addressing various geophysical problem-solving scenarios. Keywords: Semi-supervised learning, Sparse labeled data, multi-task learning, Horizon tracking, Facies classification