Interpretable ensemble meta-learning for estimated ultimate recovery prediction in shale gas wells

Accurate estimation of the Estimated Ultimate Recovery (EUR) of shale gas wells is essential for development planning and economic evaluation. However, complex geological and engineering conditions make reliable EUR prediction challenging, especially in the absence of standardized methods. Although deep learning models provide strong predictive performance, their "black-box" nature limits trust among decision-makers, and current models often neglect interpretability. To address these issues, we propose the LMTR framework, a meta-learner ensemble model for shale gas well EUR prediction. LMTR has a scaffold that integrates Lasso and Ridge regularization to enhance stability and mitigate multicollinearity, while combining a Multi-Layer Perceptron (MLP) and a Transformer to capture nonlinear relationships and long-range feature interactions. The model is evaluated using Mean Squared Error (MSE) and R², with leave-one-out cross-validation. Comparative results against 11 baseline models demonstrate that LMTR achieves superior predictive performance. In addition, SHAP analysis and Net Present Value (NPV) evaluation are incorporated to improve interpretability and support engineering optimization. The results indicate that LMTR provides accurate and reliable EUR prediction and offers a practical tool for optimizing shale gas development strategies.