Synthetic Data Augmentation to Improve Flow Zone Predictability for AI Application

Flow Zone Indicator (FZI) is a derived petrophysical parameter critical for permeability grouping and hydraulic flow unit (HFU) characterization. It aids geoscientists and reservoir engineers in delineating reservoir heterogeneity, optimizing facies modeling, and ultimately informing field development strategies [Tiab and Donaldson, 2012]. Traditional FZI calculation relies on core-derived porosity and permeability, often calibrated against well logs. However, core data is expensive and sparse, especially in brownfields, exploration wells, or tight carbonate settings. Machine learning (ML) techniques have shown significant promise in automating FZI estimation using well log inputs. Yet, ML's performance is closely tied to the availability of labeled training data. When data volume is limited or classes are imbalanced as commonly seen in reservoirs with uneven sampling across flow units ML models tend to overfit, yield unstable predictions, or fail to generalize [Zhou, 2012]. To address these challenges, we propose a robust workflow leveraging bootstrapped synthetic data augmentation. Bootstrapping, a statistical resampling technique, allows the generation of multiple training datasets by sampling with replacement. When applied carefully, it can increase the effective training sample size, capture statistical variance, and support uncertainty quantification without introducing artificial bias, provided that the underlying data distribution is respected.