Safety design is a critical concern for energetic materials, with sensitivities and performance parameters becoming increasingly important as energy density rises. However, obtaining experimental data for these properties is costly and risky. Although linear methods and neural networks have been applied to predict these properties, the limited sample size of experimental data has led to models with limitations, including inadequate accuracy, lack of effective predictive models, and inevitable reliance on experimental properties. To address these challenges, this study utilizes kernel methods and heuristic algorithms, including Genetic Algorithm and Particle Swarm Optimization, to develop effective models for predicting the impact sensitivity, electric spark sensitivity, and detonation velocity of energetic materials. After optimizing the modeling process with Particle Swarm Optimization, the models achieved R² values of 0.871, 0.898, and 0.942 on the test sets, respectively, surpassing those of neural network models, with R² values of 0.827, 0.826, and 0.909, and support vector regression models, with R² values of 0.822, 0.862, and 0.894. The proposed models significantly improve the accuracy of impact sensitivity predictions and, for the first time, offer an effective model for predicting electric spark sensitivity. By being entirely based on computational descriptors, these models expand the application range compared to previous empirical formulas. These results demonstrate the high effectiveness and accuracy of this methodology in predicting the hazardous properties of chemicals with limited experimental data.