A Dynamic Adaptive Learning Strategy for Enhanced Surrogate Modeling inThermodynamic-CFD Coupling
Please login to view abstract download link
Multiphase flow models often involve coupling a transport solver, which resolves flow equations, with a local physical model that solves thermodynamic equilibrium equations within each grid cell, subsequently computing the local thermodynamic properties of the fluid. However, the computational cost of these physical models is frequently very high. The use of surrogate models has become a standard approach to accelerate these computations. Nevertheless, significant challenges arise when the physical models depend on a large number of parameters, complicating the training process and reducing the accuracy of the surrogate models across numerous simulations. In this paper, we propose a novel, flexible, and adaptive strategy for training, validating, and inferring surrogate models. This strategy is based on the dynamic design of the dataset used for training, allowing the model to be updated and improved on-demand during simulations by incorporating new data, thereby enhancing model accuracy. Additionally, we introduce a flexible client-server architecture that enables efficient on-demand training and inference, with online prediction error estimation to ensure the quality and accuracy of the surrogate model. Our approach is validated using synthetic test cases, and we examine the impact of this active and adaptive learning method on the performance of a reservoir simulator using realistic reservoir test cases with a single injection well.
