Abstract
Two-stage light gas guns are used to accelerate projectiles to velocities of multiple kilometers per second. These guns are used in many applications including experiments to understand the response of structural components within spacecraft and satellites to impacts of orbital debris. Obtaining a precise projectile velocity can be challenging due to the many uncertainties associated with these guns. Therefore, most gas guns depend on the expertise of operators to estimate the mass of projectile, mass of main charge, the propellent gas type, the piston mass, and the pump tube (PT) pressure that would yield a specific projectile velocity. These uncertainties lead to the current need for multiple and costly gas gun experiments to achieve a certain projectile velocity.
This study aims to account for the uncertainties associated with two-stage gas guns using machine learning (ML) techniques. A dataset of 211 tests conducted at the UNLV two-stage gas gun with 0.22-inch caliber was used. Feature selection was performed and the most critical features were identified to train the ML models. The projectile velocities from each observation (experiment) were used as the dependent variable to produce training observations. Different regression techniques were evaluated. The predicted projectile velocities were then tested using the unused experimental dataset to verify the effectiveness of the models and to select the most accurate one. Among the tested models, the Random Forest model showed the best performance with R-Squared value above 94 %. The results showed that combining the experimental studies and Machine Learning can predict projectile velocities, saving time and cost of experimentation.
