Data driven solid propellant grain tailoring using surrogate optimization and ensemble machine learning

Abstract Solid propellant rocket motors (SPRMs) play a crucial role in diverse aerospace applications, with their performance intricately linked to propellant grain design. This study introduces an innovative grain tailoring approach for a highly parameterized, five-slot axisymmetric solid propellant grain. To achieve a target pressure-time history, an automated inverse design framework is developed, coupling the level set method (LSM) for robust topological burnback analysis with a well validated internal ballistic prediction module (IBPM). To investigate the most suitable data-driven strategy for inverse grain design, the study evaluates and compares two advanced data-driven optimization techniques: a Kriging-assisted Genetic Algorithm (GA) and a Random Forest Regression (RFR) machine learning model, both trained on a dataset generated via Latin Hypercube Sampling. In addition, two state-of-the-art machine learning regression models, namely XGBoost and LightGBM, are employed as benchmark models to validate the selection of Random Forest Regression for the proposed inverse design framework. While both approaches successfully navigate the 20-variable design space, the RFR model significantly outperforms the Kriging-assisted GA. Among the investigated machine learning models, Random Forest Regression consistently achieved the highest predictive accuracy and was therefore selected as the surrogate model for the proposed framework. Specifically, the RFR approach achieved a Root Mean Square Error (RMSE) of 3.24 bar compared to the Kriging model’s 11.8 bar, representing a substantial improvement in accuracy while drastically reducing computational overhead. The results demonstrate that integrating ensemble machine learning techniques like RFR overcomes the curse of dimensionality inherent in traditional surrogate models, offering a highly efficient, accurate, and automated pathway for mission-specific grain tailoring. The proposed framework therefore establishes a practical and computationally efficient methodology for high-dimensional inverse solid rocket motor grain design.