Refractory multi-principal element alloys (RMPEAs) exhibit favorable characteristics for numerous high-temperature applications. However, there is a lack of accurate approaches for tailoring the composition of RMPEAs to achieve multiple desired properties. The primary challenge results from the extensive and intricate design space that needs to be explored. Deep learning models offer a valuable advantage in extracting insights within complex systems efficiently, making them well-suited for rapidly navigating the expansive design space. In our study, we trained a deep learning framework with a prediction accuracy of 90% that is designed to predict the expected phases of given RMPEAs. Our framework is trained on a large dataset containing 50,000 RMPEAs with varying atomic percentages and elements that are labeled with their expected phases from CALPHAD calculations. However, our initial findings underscore a noteworthy decline in accuracy when applied to RMPEAs featuring elements not originally present in the training dataset. Our approach is grounded in feature distribution. Analysis of those distributions reveal that while the features of these RMPEAs align with the distributions observed in the training dataset, the correlation between features and phases for alloys is not entirely unique. To address this limitation, we introduce a generative workflow aimed at augmenting the training dataset, thereby counteracting the observed decrease in accuracy.