To evaluate the plasticization behavior of nitrocellulose， machine learning was employed with impact strength selected as the performance index. Plasticization temperature， nitrogen content， plasticization time， solvation ratio， and alcohol–ether ratio were used as independent variables to build a multi-factor quadratic regression model. Response surface methodology analyzed the main effects and interactions among these factors. Significant interaction effects are observed among the five variables. To address the limited performance of traditional linear models under small-sample and nonlinear conditions， a random forest model was combined with a nonlinear correction layer. Gaussian-noise data augmentation improved the robustness of the training set. The combined RF+GBR model achieves an R² of 0.98 and an MSE of 0.0341 （kJ·m^-2）² on the training data. Five-fold cross-validation yields an average R² of 0.95 and an MSE of 0.63 （kJ·m^-2）². These results indicate high fitting accuracy and strong generalization capability. Feature-importance analysis identifies nitrogen content as the dominant factor affecting impact strength， followed by solvation ratio. The study provides a quantitative basis for evaluating plasticization reliability and optimizing process parameters.