Sample-aware regression with selective generative augmentation and graph learning for aircraft engine weight prediction under data scarcity

Accurate prediction of aircraft engine weight during the early design phase is crucial for aircraft performance evaluation and design decision-making. However, this task remains challenging in resource-constrained engineering settings because available data are often scarce, imbalanced, and characterized by complex interactions among design parameters. To address these challenges, this study proposes a sample-aware regression framework with selective generative data augmentation and graph-based feature learning (SAR-SAGL). First, a selective generative augmentation module identifies hard-to-fit regions and synthesizes local samples to mitigate data scarcity and distribution imbalance. Second, a feature correlation graph is constructed to represent interdependencies among engine design parameters, and a Graph Convolutional Network is used to derive graph-enhanced structural embeddings. Finally, a sample-aware regression mechanism dynamically adjusts sample weights according to fitting errors, enabling the model to focus more on difficult samples during training. Experiments on real-world datasets from NASA and commercial aero-engine projects demonstrate that SAR-SAGL achieves the best overall performance among all compared models. Additional validation and diagnostic analyses further support the effectiveness and applicability of the framework. Overall, this work establishes a promising pathway for sample-efficient regression in data-constrained engineering design.