Large-scale water resources optimization often involves using time-consuming simulation models to evaluate potential water resource designs or calibrate parameter values. Approximation models have been proposed for improving computational efficiency of the optimization. In most instances, multiple simulation runs have been done prior to the optimization, which are then used to fit an approximate model that is used during the optimization. This paper demonstrates that this approach can lead to suboptimal solutions and proposes a dynamic modeling approach, called Adaptive Neural Network Genetic Algorithm (ANGA), in which artificial neural networks are adaptively and automatically trained directly within a genetic algorithm (GA) to replace the time-consuming water resource simulation models. A dynamic learning approach is proposed to periodically sample new solutions both to update the ANNs and to correct the GA's convergence. Different configurations of ANGA were tested on a hypothetical groundwater remediation design case, and then the best configuration was applied to a field-scale case. In these applications, ANGA saved 85--90% percent of the simulation model calls with no loss in accuracy of the optimal solutions. These results show that the method has substantial promise for reducing computational effort associated with large-scale water resources optimization.