Abstract—Using a solver, e.g. genetic algorithm (GA), to optimize over a numerical CFD simulator is computationally expensive. Hence, surrogate-based optimization (SBO), is a common approach to efficiently find the optimal solution. Machine learning (ML) techniques hold enormous potential in building models that provide accurate, fast, and computationally inexpensive proxies that feed the SBO loop and map the objective function with minimal sample points, allowing the optimization solver to locate the global optimum. We demonstrate the feasibility of this by succesfully modeling a physics-based numerical reservoir CFD simulator using ML that achieves comparable accuracy while substantially reducing the computational cost.
Full paper here.