A precise characterization of permeability in low-permeability oil and gas fields is a pressing issue in petroleum engineering, a field in which standard well tests are ineffective. Diagnostic Fracture Injection Tests (DFITs) have become important diagnostic characterization tools, but existing after-closure analytical tools have severe limitations, such as sensitivity to closure identification and limited applicability to pseudo-radial flow regimes. Although machine learning uses for DFIT interpretation have been investigated, available literature is mainly based on synthetic data, is less concerned with predicting permeability and more with closure pressure, and lacks rigorous interpretability analysis to be accepted by industry. In this research, a collection of machine learning models, which enable prediction of reservoir permeability given the parameters of DFIT G-functions, is developed and verified on the basis of a large field database comprising 620 tests. Ten algorithms were compared in a systematic fashion, which comprised neural networks, gradient boosting algorithms (CatBoost, XGBoost, LightGBM), regularized linear models, and ensemble algorithms. Multi-Layer Perceptron Regressor was more efficient with R² = 0.9576, RMSE = 0.0357 md, and MAE = 0.0257 md on training data with good generalization on blind validation (155 independent tests: R² = 0.9194, RMSE = 0.0805 md). SHAP analysis established physically significant relationships, with G-time at closure having the most significant negative correlation with permeability. The symbolic regression using genetic programming gave a complementary equation-based model with an R² of 0.9379. The resulting models are much more efficient than traditional analytical methods and offer quick, precise, and interpretable forecasts that can be used to optimize hydraulic fracturing designs and plan field development in challenging reservoir conditions.
Nashed et al. (2026) studied this question.