Physics-Constrained Production Forecasting with Direct Fracture Characterization by Field Data

Abstract Fractured reservoir simulations often rely on abstract fracture parameters—such as length, height, and width—that are difficult to validate with direct measurements, introducing biases in production forecasting. This research presents Assimilation Neural Network (AssimNet), a hybrid data-driven and simulation-proxy framework that completes the history-matching process during training. By directly integrating real-world measurements (e.g., wellhead, completion, logging data) into simulation inputs, AssimNet ensures direct characterization from field data and provides physics-constrained production forecasting. AssimNet leverages a simulation proxy to capture the underlying physics of numerical models. It encodes simulation inputs and the field measurements into the shared latent space, and a shared decoder reconstructs simulation inputs from the field-derived latent representation, ensuring physically consistent and interpretable outputs. Evaluated on 1,488 wells from the Montney Shale, AssimNet consistently outperforms traditional data-driven models in data-sparse environments. AssimNet is the first hybrid model that directly translates field measurements into physics- consistent simulation parameters optimized for the given simulator. It offers a more transparent, efficient, and interpretable approach to reservoir characterization and production forecasting, enabling faster, data-driven decisions in subsurface engineering. Unlike traditional physics-informed neural networks or neural operators, AssimNet tightly couples production data with numerical simulation, minimizing reliance on abstract assumptions and iterative tuning. By embedding simulation within a data-driven architecture, it enhances generalization, improves accuracy in data-sparse regions, and reduces uncertainty in fractured reservoir forecasting.