Abstract
<title>Abstract</title> <p>Waterflooding well-control optimization seeks the time-dependent injector rates and producer bottom-hole pressures that maximize the Net Present Value (NPV) under well- and field-level constraints. Solving it requires thousands of full-physics reservoir simulations, each of which may take hours, making direct simulator-in-the-loop optimization impractical at field scale. Machine-learning proxy models can replace the simulator within the optimization loop, but most studies assess them only by prediction accuracy, leaving open whether the resulting control strategies remain effective when re-evaluated in the underlying high-fidelity simulator. This work presents a simulator-validated engineering workflow that transforms a Long Short-Term Memory (LSTM) controlled production time-series proxy into a reliable decision-support tool. The workflow integrates Bayesian Optimization with the Tree-structured Parzen Estimator (BO-TPE) for automatic hyperparameter selection, a Genetic Algorithm (GA) for NPV maximization, and an NPV-aware smoothing regularization that discourages abrupt cycle-to-cycle control variations. It is evaluated on two reservoirs of increasing complexity: a synthetic two-phase model and the Brugge benchmark. Systematic sensitivity analyses quantify the influence of the hyperparameter budget, random-startup fraction, smoothing factor, and GA population size. Every optimized solution is re-evaluated in a high-fidelity simulator, where a top-K re-ranking identifies the simulator-confirmed best strategy while revealing the residual proxy bias. The results show that the most accurate proxy does not necessarily yield the highest economic return, highlighting that predictive accuracy alone is insufficient to guarantee optimization quality. Overall, the proposed workflow provides a reproducible, rigorously validated framework for transforming predictive machine-learning proxies into trustworthy optimization and decision-support tools for practical reservoir management.</p>