Support-vector regression accelerated well location optimization: algorithm, validation, and field testing.

We have developed a machine-learning (ML) accelerated optimization method for challenging real-life well-location optimization (WLO) applications. The optimization protocol encompasses a stage at which the objective function is computed by a proxy ML model trained to accurately approximate the objective function that would be otherwise calculated using a computationally more intensive reservoir simulation run. As ML algorithms for proxy development, we have used two variants of the support-vector regression (SVR) technique, namely, ε-SVR and ν-SVR. In the proposed proxy-accelerated optimization workflow, WLO starts with the conventional approach of using the reservoir simulator as the objective-function evaluator for a relatively small number of iterations of a global optimizer that conducts explorative search. This initial simulation investment is utilized not only to deliver optimization improvements but also to train an SVR proxy for the objective function of the investigated WLO problem. Subsequent to initial iterations, the SVR proxy is activated to serve as the objective-function evaluator for the global optimizer. Upon convergence, optimal model parameters estimated by the proxy-based optimization protocol are input to the flow simulator to access the full set of simulation results and also to re-evaluate the proxy accuracy. Optionally, the optimal model configuration can be supplied as initial guess to an efficient local optimizer to evaluate the further improvement potential for a few iterations using the reservoir simulator as the objective-function evaluator. The proposed SVR-accelerated optimization method is tested on a real field WLO problem. We have observed a 40 to 60% reduction in the overall computational cost with the SVR-accelerated protocol. It is important to note that the training time of the SVR proxy is negligibly short for small datasets (with 100s to 1000s data points) such as the ones generated over the course of WLO iterations. However, the main challenge associated with the proposed method is tuning the hyper-parameters of the SVR proxy for a given problem of interest and striking a good balance between accuracy and generalizability of proxy predictions. While some experimentation is still required with SVR hyper-parameters, especially in the first application, we accelerate this step through an auditable optimization process in our implementation. Results of our study demonstrate that the SVR-accelerated WLO method is applicable for real-life problems. [ABSTRACT FROM AUTHOR]