Interactive ML-Assisted Seismic Interpretation Using Advanced Meta-Labelling

This paper presents an innovative workflow for seismic interpretation, addressing the time-consuming task of manually interpreting geological horizons and geo-bodies in complex areas. Traditional machine learning approaches face challenges with initial labeling, where the trained models are not general enough to be applicable on the entire seismic data. This method overcomes through an incremental development of labels, starting from areas with certain geological segmentation and progressively expanding to less evident areas. This approach, grounded in meta-learning and meta-labeling, enables the machine to automatically adjust itself through meta-learning process. The methodology involves training a deep learning model concurrently with label expansion. In the beginning, the model adapts to local seismic data characteristics, using very limited initial labels (e.g., from well trajectories or a pseudo well) to generate and refine predictions in adjacent areas. This iterative process of prediction, evaluation, correction, and retraining continues until satisfactory results are achieved. The effectiveness of this workflow is demonstrated in its application to a real 3D seismic dataset, showing the potential of the meta-learning process for automatic seismic interpretation.