Behnam Jafarpour

University of Southern California

Active learning for efficient well control optimization

Fit-for-purpose surrogate models are used to speed up the solution of well control optimization by reducing the number of full-scale reservoir simulation runs. Machine learning (ML)-based surrogate models rely on extensive amount of simulated data during off-line/passive training. Once trained, they provide fast predictions that can be used for gradient-based well control optimization. However, off-line training can be inefficient and lead to predominantly extrapolation beyond the training data, and with limited accuracy.  In this presentation, I present active learning to adapt the training to the optimization path to preserve the local accuracy of the model during gradient-based optimization. As the optimization progresses, data samples that are less relevant are replaced with new local samples to help improve the prediction power of the surrogate model along the optimization path. I present several numerical experiments to compare the performance of active learning approach with that of the traditional off-line training framework.