Physics‑decomposed residual learning with PolyRF‑boost for cold gas thrust prediction.

Area of Science:

• Aerospace Engineering
• Computational Chemistry
• Propulsion Systems

Background:

• Accurate prediction of cold gas propulsion thrust is crucial for spacecraft design and mission planning.
• Existing methods may struggle to capture the complex interplay between macroscopic physical behavior and microscopic chemical deviations.

Purpose of the Study:

• To develop and validate a new computational method, PolyRF-Boost (Physics-Decomposed Residual Learning - PDRL), for precise average thrust prediction in cold gas propulsion.
• To demonstrate the model's effectiveness in propellant ranking and aerospace system design.

Main Methods:

• The PolyRF-Boost method decomposes the target function into a macroscopic physical component (Backbone Polynomial Ridge) and a microscopic chemical deviation component (Gradient Boosting residual).
• Data preprocessing involved feature engineering, selection (Random Forest), and outlier mitigation using an absolute value error function.
• Model performance was evaluated using 5-fold cross-validation with MAE, RMSE, and R2 metrics.

Main Results:

• PolyRF-Boost/PDRL achieved a test R2 of 0.9899, outperforming baseline models like Polynomial Regression (R2=0.9887), Random Forest (R2=0.8056), and Gradient Boosting (R2=0.7667).
• The model demonstrated strong generalization on real data, with test set R2 exceeding 0.98.
• Theoretical analysis supported the efficacy of separating physical and chemical components for improved generalization bounds.

Conclusions:

• The PolyRF-Boost/PDRL method accurately predicts cold gas propulsion thrust by effectively modeling both physical and chemical aspects.
• This approach shows significant potential for optimizing propellant selection and enhancing aerospace system design.
• The decomposition strategy offers a theoretical advantage over integrated modeling techniques.