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

Abstract

Abstract In this paper, a method of compilation and leveling is provided to accurately predict the average thrust of cold gas propulsion. The proposed method, which is described as PolyRF-Boost and implemented in the form of conceptual framework Physics-Decomposed Residual Learning (PDRL), explicitly divides the target function into two independent components: a physical macroscopic process (( x )) that describes ideal behavior and a microscopic chemical deviation (( x )) that corrects reality gaps. Accordingly, the final estimator is defined as the sum total of two layers as $$\hat{\user2{f}}\left( {\varvec{x}} \right) = {\mathbf{\mathcal{M}}}_{{{\varvec{\theta}}_{{{\varvec{phys}}}} }} \left( {\phi \left( {\varvec{x}} \right)} \right) + {\mathbf{\mathcal{G}}}_{{{\varvec{\theta}}_{{{\varvec{resid}}}} }} \left( {\varvec{x}} \right)$$ , in which () the macroscopic layer is based on a Backbone Polynomial Ridge on physically engineered features (such as $${\varvec{I}}_{{{\varvec{sp}}}}^{2}$$ and $${\varvec{P}}_{{\varvec{T}}} /{\varvec{I}}_{{{\varvec{sp}}}}$$ ) and $$\phi \left( {\varvec{x}} \right)$$ is second-degree polynomial conversions of those characteristics, and is a refining residual based on Gradient Boosting, which models the complex chemical deviations. The dataset used was loaded from the Excel file Propellant_Ranking_FP.xlsx ( http://dx.doi.org/10.5281/zenodo.7765215 , version 1.0.0, released March 24, 2023). After sampling 500 rows and preprocessing, the final dataset contained 500 samples with 21 features. Preprocessing included detecting and removing data leakage (removing 2 features: Total_Impulse and Average_Satellite_Mass due to high correlation), removing 7 identifier columns and 11 redundant correlated features. Then, polynomial feature generation, feature selection with Random Forest (top-k selection; in practice top-12) and correction with Gradient Boosting were performed. To increase noise robustness, the residual layer was trained with an absolute value error function to reduce sensitivity to outliers. Evaluation with 5-fold cross-validation and MAE, RMSE and ( R 2) measures shows that PolyRF-Boost/PDRL achieves competitive performance with a test R2 of 0.9899, MAE of 0.000312, and MSE of 0.000006, outperforming baseline models including Polynomial Regression (R2 = 0.9887), Random Forest (R2 = 0.8056), and Gradient Boosting (R2 = 0.7667); actual vs. prediction and error comparison plots confirm this performance. Experiments on real data show that ( R 2) in the test set exceeds 0.98, which demonstrates the model’s ability to learn the complex physicochemical structure of cold gas propulsion systems and its applicability to propellant ranking and aerospace system design. Finally, a theoretical analysis is also presented—including a decomposition error bound theorem—that shows that separating the physical process (Bias reduced via Ridge) and the dense residual learning (Exponential error reduction with boosting) leads to a tighter generalization bound than the integrated approaches.