MGPNet: A multi-modal geo-physical production network for reservoir yield forecasting

Abstract

Accurate reservoir production prediction often depends on a single data source or simplified physical assumptions, limiting the ability to fully capture complex reservoir heterogeneity and multi-physical coupling processes. To address these challenges, this paper proposes MGPNet, a multi-modal reservoir production prediction network that integrates seismic images, logging curves, and historical production sequences. The network incorporates the GeoFE-AE module, which uses adversarial feature enhancement and semantic alignment mechanisms to achieve deep feature extraction and collaborative representation across modalities. Furthermore, it introduces a Multi-scale Cross-Attention Mechanism (MCAM) to enable effective feature fusion across different modalities at multiple semantic granularities, and employs a bidirectional Transformer decoding and prediction module (BiAD) to accurately forecast future production sequences. Using a high-fidelity synthetic multi-modal dataset, experimental results demonstrate that MGPNet significantly outperforms existing mainstream methods across several key metrics: Mean Absolute Error (MAE) of 2.12, Root Mean Square Error (RMSE) of 5.87, Coefficient of Determination (R²) of 0.987, and Explained Variance Score (EVS) of 0.971. These results validate the model’s comprehensive strengths in accuracy, stability, and robustness to noise. Furthermore, transfer-learning evaluations on real wells from the Volve oilfield confirm the model’s practical applicability and strong cross-well generalization capability. This research offers a promising technical approach for deep fusion modeling of multi-source reservoir data, with substantial potential for practical engineering applications and further academic exploration.