Accurate prediction of physical stability in amorphous solid dispersions (ASDs) is essential for the design of successful pharmaceutical formulations. However, small datasets and class imbalance are still significant challenges that result in models suffering from multiple problems. In this work, we propose an enhanced deep learning model for predicting ASD physical stability while addressing class imbalance. The approach uses a hybrid sampling strategy that combines SMOTE and Tomek Links to address disbalance through both the oversampling of the minority class and the removal of ambiguous boundary instances. Autoencoders are used for dimensionality reduction for nonlinear data relationships and reducing computational complexity. Class weights are adjusted within the loss function during training to prioritize accurate predictions for the minority class. Experimental results highlight the model’s superior accuracy, precision, recall, faster convergence, and reduced computational cost compared to traditional approaches. This framework demonstrates the potential of advanced deep learning techniques to enhance pharmaceutical modeling and accelerate drug development.