Interpretable Ensemble Meta-Learning for Supervised Prediction of Estimated Ultimate Recovery in Shale Gas Wells

Accurately assessing the Estimated Ultimate Recovery (EUR) of shale gas wells is crucial for informed decision-making in development planning and maximizing economic benefits. However, the complex geological conditions and engineering schemes involved create a unique and intricate flow environment, making EUR highly sensitive to the interplay of geological, engineering, and production factors. Given the complexity of flow mechanisms and the lack of a clear, standardized calculation method, reliably predicting EUR remains a challenge. While deep learning models offer substantial predictive power, their "black-box" nature limits trust among decision-makers, and current models often neglect interpretability. To address these challenges, we introduce the LMTR framework, a meta-learner ensemble model tailored for EUR prediction in shale gas wells. LMTR integrates Lasso and Ridge regularization at its two ends, effectively mitigating multicollinearity and ensuring a stable linear structure. This scaffold supports a multi-layered architecture, where a Multi-Layer Perceptron (MLP) captures nonlinear relationships, and a Transformer component models long-range feature interactions. This hybrid design strikes a balance between interpretability and predictive capability, offering a robust framework for EUR estimation.The model's performance was evaluated using Mean Squared Error (MSE) and R² metrics, with cross-validation conducted via a leave-one-out method. Comparative experiments with 11 baseline models demonstrate superior predictive accuracy of LMTR, suggesting its potential as a decision-support tool for early-stage shale gas development. Additionally, the framework enhances interpretability through SHAP analysis and integrates Net Present Value (NPV) calculations, guiding engineering optimization based on geological variability. The results indicate that LMTR outperforms existing approaches in EUR prediction and provides a reliable basis for optimizing development strategies.