An interpretable deep learning framework using FCT-SMOTE and BO-TabNet algorithms for reservoir water sensitivity damage prediction

This study proposes an interpretable deep learning framework to address the high-dimensional and inherently unpredictable challenges associated with oil and gas drilling and completion operations. By comparing TabNet, Tab Transformer, Hopular, and TabDDPM through computational experiments under identical conditions, TabNet was selected as the optimal approach. The framework integrates Bayesian optimization (BO) with TabNet to model complex oilfield tabular datasets. Fair Cut Tree (FCT) and Synthetic Minority Over-sampling Technique (SMOTE) are incorporated to mitigate data missingness and imbalance, thereby enhancing dataset integrity and robustness. Empirical validation was conducted using 270 data entries collected from 15 distinct oil fields, specifically focusing on reservoir water sensitivity damage in natural core samples. The proposed framework exhibited superior predictive accuracy for the water sensitivity index on an independent test set, achieving a mean absolute percentage error (MAPE) of 4.4495% and a root mean square error (RMSE) of 4.05, underscoring its strong generalization capability. Moreover, this methodological approach enables a quantitative assessment of the influence of critical factors, including reservoir water salinity, initial permeability, and the mineralogical composition of rock formations, on water sensitivity predictions. This represents a significant advancement from traditional qualitative analyses to a more rigorous quantitative factor analysis, with the interpretability findings corroborating established mechanistic insights. The proposed framework offers a versatile and reliable solution for precise predictive modeling in complex drilling and completion scenarios reliant on tabular data, thereby providing a robust theoretical foundation and algorithmic support for accurate forecasting in the oil and gas industry.