Empirical Models for Predicting Two-Stage Light Gas Gun Muzzle Velocity

Two-stage light gas guns (2SLGGs) can reliably accelerate projectiles to velocities between 1.5 and 8.0+ km/s and are used in hypervelocity impact, aerospace, and hypersonic research. 2SLGG operation involves a variety of physical phenomena including combustion, gas-compression, heat-transfer, and friction. Due to the wide range of operational parameters and experimental uncertainty, accurate muzzle velocity predictions can be a serious challenge. In this paper, a series of regression models for predicting muzzle velocity were fitted to and validated against 171 2SLGG launches (projectile velocities, 1.5–6.8 km/s) performed at the Texas A&M University Hypervelocity Impact Laboratory. Most of the regression models analyzed had minimal accuracy improvement compared to basic linear regression. However, a neural network model (RMSE = 0.234 km/s) utilizing several methods to combat over-fitting, showed consistent improvement over linear regression (RMSE = 0.260 km/s) and Gaussian process regression (RMSE = 0.240 km/s). Regression model projectile velocity estimates were compared to results from the classical Piston Compression Light Gas Gun Performance (PCLGGP) and state-of-the-art LGGUN 2SLGG performance prediction codes. All of the regression models demonstrated significantly better predictive capabilities than the PCLGGP model (RMSE = 0.597 km/s), particularly at lower velocities. The regression model absolute errors from the 2SLGG experiments also compared very favorably to general absolute error estimates obtained using LGGUN. These results suggest that easy-to-implement, maintain, and scalable regression models may provide a viable alternative to complex physics-based computational models for 2SLGG launch velocity predictions, particularly as the volume of available experimental data increases. Such regression models have the potential to markedly improve predictive capabilities, identify complex coupling between experimental parameters, and reduce uncertainty.