Accurately estimating aircraft fuel flow is critical for evaluating new procedures, designing next-generation aircraft, and monitoring the environmental impact of current aviation practices. This paper investigates the generalization capabilities of deep learning models for fuel flow prediction, focusing on their performance with aircraft types not included in the training data. We propose a novel methodology that combines neural network architectures with domain generalization techniques to improve robustness and reliability across different aircraft types. Using a comprehensive dataset of 101 aircraft types, split into training (64 types) and generalization (37 types) sets with each type represented by 1,000 flights, we introduce a pseudo-distance metric to quantify aircraft type similarity and explore sampling strategies to improve model performance in data-limited regions. Our findings show that for unseen aircraft types, especially with noise regularization, the model outperforms baselines such as corrected proxy estimates. This study demonstrates the potential of blending domain-specific insights with advanced machine learning techniques to develop scalable, accurate, and generalizable fuel flow estimation models.