Decision-making under uncertainty addresses real-world problems that are often hard to solve, particularly when involving both discrete and binary decision variables. In two-stage stochastic programming problems, decisions are made before uncertain data is revealed, with adjustments optimized in a recourse problem after the revelation. The ideal solution is robust and resilient across all possible future scenarios. Classic solving techniques rely on selected sets of scenarios to approximate the problem. Accuracy improves with the inclusion of more scenarios and concurrently, the complexity of the problem increases, frequently resulting in instances that are computationally intractable. With the rapid advancement in machine learning technologies, our study aims to train neural networks as surrogate models for the recourse problem. These networks aim to predict solution quality across diverse scenarios, thereby reducing the complexity of the original optimization model. Through empirical analysis, we present preliminary findings demonstrating the application of this predictive capacity and examine the practical challenges encountered.