Combinatorial problems are among the most complex challenges encountered in real-world scenarios. Examples include task scheduling, supply chain logistics, and optimal product packing. Adaptive operator selection algorithms, a class of artificial intelligence algorithms, autonomously select and apply the most effective strategies to solve such problems. Specifically, these algorithms choose and apply a heuristic at each step to improve a current solution. These algorithms require two components: Credit Assignment, which associates a heuristic with a reward value, and Selection Rule, which selects the heuristic to be applied based on its assigned reward. The process of credit assignment is computationally intensive, as it typically involves evaluating the potential success of a given heuristic within the current solution. This computational burden limits the overall efficiency of operator selection algorithms. In this study, we propose the use of surrogate models. In particular, the implementation and training of a Feedforward Artificial Neural Network is proposed, which will be trained as a metric for evaluating evolvability potential. Surrogate models provide approximations to functions that, while not perfectly accurate, are computationally easier and faster to evaluate than the original models. This novel approach will be tested on well-established instances of combinatorial problems such as Onemax and TSP.