Improving deep neural network prediction of hydraulic fracturing initiation pressure in hydrate-bearing sediments through feature engineering and architecture search

Abstract Accurate prediction of hydraulic fracturing initiation pressure ( $$P_b$$ ) in natural gas hydrate-bearing sediments provides screening-level support for laboratory-scale hydraulic fracturing design in NGH-bearing sediments. This study demonstrates that systematic feature engineering and neural architecture search yield substantial improvements for predicting $$P_b$$ on a 114-sample published dataset. The framework first expands the feature space from 12 to 208 variables via polynomial transformations, interaction terms, and domain-specific features, followed by deduplication. Multi-method feature selection identifies a 14-feature subset, and extended evaluation identifies a verified 12 $$\times $$ 32 model as the best-performing saved configuration. The reproducible-by-retraining reference is R $$^2$$ $$\approx $$ 0.87; the verified 12 $$\times $$ 32 saved model achieves R $$^2$$ = 0.92 on its original held-out split and is reported as an upper bound on the dataset. The best of 60 retraining runs reaches 0.85—none reproduces 0.92, confirming that 0.92 is a saved-artifact result, not a reproducible target. The verified 12 $$\times $$ 32 model reduces MSE from 19.34 to 8.22 MPa $$^2$$ and MAE from 3.16 to 2.14 MPa compared to the best prior reported result (R $$^2$$ = 0.81). Feature importance and SHAP analysis show that the model assigns highest importance to stress-composition interactions and stress-derived statistical terms, consistent with known geomechanical controls on fracture initiation. Generalization is modest (CV mean R $$^2 \approx $$ 0.50), and simpler models (RandomForest, SVR) are recommended as production-ready baselines; the 12 $$\times $$ 32 DNN is retained as the best verified saved artifact. These findings confirm that the dataset contains untapped predictive capacity, and that methodological refinement through domain-informed feature engineering combined with systematic architecture search provides an effective pathway for improving prediction fidelity in small-sample geomechanical problems.