Perfecting the Data, Not the Model: Interactive Multi-Volume Deep Learning for Complex 3D Geobody Interpretations

Objective / Scope This paper introduces a data-centric deep learning (DL) workflow that focuses on improving seismic data quality rather than tuning model hyperparameters. Using sand injectites from the Volve Field in the Norwegian North Sea as an example, the workflow demonstrates that prioritizing data quality over model selection and hyperparameter tuning improves the accuracy, consistency, and interpretability of 3D geobody network predictions. The objective is to establish a repeatable, geologically grounded framework that minimizes model bias by emphasizing data integrity and its relationship to geologic truth.