Supersonic Vehicle Zero-Lift Drag Coefficient Identification via Transformer-Based Direct Mapping
Year: 2026
Type: conference-paper
Abstract: Accurate identification of the supersonic vehicle zero-lift drag coefficient Cd0(Ma) is fundamental for trajectory prediction and flight state estimation. Traditional methods rely on explicit dynamics models and iterative optimization, suffering from model mismatch, high computational cost, and sensitivity to initial values. This paper proposes a Transformer-based direct mapping framework that outputs the entire Cd0(Ma) curve in a single forward pass from a trajectory sequence, without iterative optimization or dynamics equation integration. The model consists of a trajectory Transformer encoder, a physics extractor, a Mach embedding module, and a 3-layer MLP direct mapping layer. The model is trained with Huber loss and second-order difference smoothness regularization. On a dataset of 3,456 trajectories, the method achieves R²=0.8157, MAE=0.0240, and RMSE=0.0284 on the test set. Under the assisted calibration setting—where a single external reference value Cd0(Ma_ref) is available (e.g., from wind tunnel experiments or CFD)—a vertical shift post-processing correction further improves R² to approximately 0.99, reducing MAE by 79% and RMSE by 75%. This result represents an externally-assisted calibration scenario and should be distinguished from the end-to-end inference performance. Ablation studies validate the effectiveness of each module, with Mach embedding identified as the most critical component.
Institution: Nanjing University of Science and Technology
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204759473