Insights on Shale Gas Production Prediction via Combining Multimodal Machine Learning with Deterministic Geological Modeling

Traditional machine learning models for shale gas production prediction usually ignore spatial reservoir heterogeneity. This paper develops a multimodal machine learning (MML) method integrating a deterministic geological model with multi-source petrophysical data. The model combines a CNN for spatial feature extraction and an ANN for tabular data analysis, significantly improving prediction accuracy. It is found that the MML model achieves a superior coefficient of determination (R2 = 0.849) for 12-month Duvernay shale gas productivity forecasting, significantly outperforming the standalone ANN model (R2 = 0.719). Furthermore, the MML approach successfully explains production discrepancies between horizontal wells with similar average reservoir properties—a challenge traditional ANN models fail to address—by capturing lateral formation variability. Validation on three representative wells confirms high prediction accuracy, with conformity rates of 89.1%, 91.8%, and 92.6%, respectively. This method not only enhances production forecasting but also provides a practical tool for optimizing horizontal well placement, enabling more efficient development of shale gas reservoirs. The study highlights the potential of multimodal machine learning in unconventional resource exploitation, providing a data-driven approach to reservoir management.