A Physics-Guided Gradient Boosting Framework for Open-Source Aviation Fuel Estimation

Abstract

Accurate segment-level fuel consumption estimation is essential for aviation emission accounting and cost management, yet existing models rely on proprietary flight recorder data and fail to generalize across heterogeneous open trajectory sources. This paper presents a physics-guided gradient boosting framework for direct fuel consumption prediction, solving the EUROCONTROL PRC 2025 Data Challenge. The pipeline integrates raw ADS-B trajectories with METAR meteorological observations, airport infrastructure data, aircraft performance coefficients, and regional passenger load factors, including segment boundary injection, linear interpolation, and timestamp correction. Physics-informed features are derived by running the OpenAP FuelFlow model with dynamic mass tracking over complete flight trajectories, producing physically consistent fuel burn estimates for all 36 aircraft types in the dataset. To address the narrowbody/widebody class imbalance, synthetic widebody training samples are generated via Gaussian noise perturbation with kinematic bounding, increasing the widebody training share from 18.7% to 34.1%. Sequential Forward Selection reduces 133 candidate features to a compact subset, and Bayesian hyperparameter optimization via Optuna identifies the final XGBoost configuration. On the held-out validation set, the proposed model achieves a mean absolute error (MAE) of 85.7 kg, RMSE of 204.6 kg, and R² = 0.961 over segments with an average duration of 8.87 minutes. On the official hidden competition test sets, which have an average segment duration of 7.24 minutes, the model achieves an RMSE of 214.92 kg, outperforming standard machine learning baselines such as Random Forest and LightGBM. Comparison against linear models such as Ridge Regression (MAE 262.2 kg, R² = 0.355) confirms the high nonlinearity of the task. These results demonstrate that competitive fuel consumption estimation accuracy is achievable from publicly available data alone, without reliance on proprietary airline records.