Hybrid RGNet: A physics-based model enhanced by machine learning for history matching, characterization and future prediction

As a reduced-physics model, RGNet (reservoir graph network) represents a reservoir model that captures essential features of flow in petroleum reservoirs while simplifying complexities. RGNet has been successfully applied to various reservoir management problems, spanning reservoir connectivity analysis, resource volume estimation, forecasting, and flood optimization. To further improve the model accuracy and predictability, we develop a hybrid version of RGNet which is enhanced by machine learning (ML). We propose to use an ML-enhanced history-matching algorithm (SVR-DGN) to identify multiple best-matched realizations for RGNet. The ML model is modified to simultaneously learn multiple outputs, which significantly reduces the training cost. Additionally, we propose a sparsity-constrained history-matching algorithm and a physics-informed machine learning (PIML) method. With these methods, we can obtain well connectivity in a minimum viable pattern and other model parameters while simultaneously matching the data and enhancing the forecasting accuracy of RGNet models. The proposed workflow was applied to both a synthetic and a deepwater field example. In the synthetic example, the sparsity-constrained history-matching successfully identified key inter-well connections while matching the data. In contrast to alternate methods that pre-determine well connections, our proposed approach considers the uncertainty in well-connection patterns and provides multiple realizations for connecting wells. In the field example, SVR-DGN is shown to simultaneously match multiple types of observed data, including bottom-hole pressure (BHP), gas-oil ratio (GOR), and water-cut, with good blind-testing quality. We also tested whether PIML can further improve RGNet accuracy for BHP and water-cut forecasts in the field example. The best-matched model was used as the base-case scenario, and its matching error was taken as the training output for the PIML model. During the forecast period, we applied the PIML model as the “corrector” to the RGNet responses. The results showed that the PIML-corrected BHP demonstrated much better agreement with the real data compared to the original RGNet responses. In reservoir dynamic modeling, RGNet can capture the major features of a target reservoir while potentially overlooking minor details due to “reduced” physics. The integration of ML into RGNet adds more model details for small features, thereby improving model accuracy. The proposed workflow enhances the capabilities of RGNet by improving model fidelity without compromising its rapid computational speed. • Hybrid approach for reservoir modeling. • ML-enhanced history matching. • Sparsity-constrained reservoir connectivity. • PIML for improving forecast.