Multiphase Flow Simulation in Hydrocarbon Reservoir: Balancing Computational Efficiency and Prediction Accuracy

Résumé fourni par la source

ABSTRACT High‑fidelity reservoir simulations remain indispensable for predicting multiphase flow in porous media, yet their computational cost impedes real‑time decision‑making. Recent advances have demonstrated that deep‑learning proxies—such as convolutional and recurrent neural networks—can capture complex flow dynamics, albeit often at the expense of substantial model complexity and training overhead. In this context, we systematically evaluate whether simpler, more transparent machine‑learning (ML) regressors can serve as efficient and reliable surrogates, achieving competitive accuracy with far lower computational demands. Using synthetic datasets generated with the MATLAB Reservoir Simulation Toolbox (MRST) for both single‑phase and two‑phase heterogeneous reservoirs, we benchmark four algorithms: k ‑nearest neighbors ( K NN), random forests, decision trees, and artificial neural networks (ANNs). Among these, the K NN regressor attains the highest predictive performance ( R 2 = 0.95, root mean squared error [RMSE] ≈ $8.1 million for net present value) while requiring only 0.004 s per iteration—over two orders of magnitude faster than conventional full‑physics runs. Random forests and ANNs also capture nonlinearities effectively, but K NN offers the best accuracy‑efficiency trade‑off. Our results confirm that carefully trained, simplicity‑driven proxies can rival state‑of‑the‑art deep‑learning models in predictive accuracy, yet with markedly lower complexity and computational cost. The contribution lies not in proposing novel algorithms, but in providing a rigorous comparative assessment that highlights the practical viability of lightweight ML surrogates for rapid production forecasting, injector placement optimization, and economic screening. Future work will focus on integrating physics‑informed constraints, uncertainty quantification, and scalable validation on field‑scale data to further strengthen their applicability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.