Category
Applied
Description
In rocketry, small variations in weather and launch conditions can cause large differences in the highest altitude (“apogee”) a rocket can reach during flight, which makes targeting a precise altitude difficult. Airbrakes—small and deployable surfaces that increase drag—can actively adjust a rocket’s climb to achieve the intended apogee. Our goal is to develop an airbrake deployment approach that can translate to real rocket applications for teams lacking resources such as high-fidelity computational fluid dynamics, wind tunnel testing, or multiple test launches. Therefore, rather than using a control method that continuously adjusts the amount the airbrakes are deployed into the airstream, we limit our control methodology to determine a single deployment altitude with full deployment, after which the airbrakes remain deployed. If the rocket begins to undershoot the target altitude, the airbrakes retract early and remain retracted. This research investigates how various supervised machine-learning (ML) models can select the airbrake deployment altitude so that a simulated rocket reaches a 10,000-foot target apogee within one percent error. This simulation-based, quantitative, single-case design uses RocketPy, an aerospace simulation tool, to generate a large set of simulated flights with different airbrake deployment altitudes under varying environmental and launch conditions. This dataset is used to train random forest, gradient boosting, and support vector regression models to predict the optimal deployment altitude. Each model is evaluated using new RocketPy simulations with different flight conditions to compare how accurately each model achieves the target apogee. Our research question is: “Which ML model most effectively minimizes apogee error relative to the 10,000-foot target for the Calisto rocket when trained and tested through RocketPy simulations?” Model success is defined as error below one percent. This study supports future work developing more advanced ML-based controllers for physical rocket systems under site-specific weather and launch variations.
Machine-Learning based Airbrakes Control for Rocket Apogee Targeting
Applied
In rocketry, small variations in weather and launch conditions can cause large differences in the highest altitude (“apogee”) a rocket can reach during flight, which makes targeting a precise altitude difficult. Airbrakes—small and deployable surfaces that increase drag—can actively adjust a rocket’s climb to achieve the intended apogee. Our goal is to develop an airbrake deployment approach that can translate to real rocket applications for teams lacking resources such as high-fidelity computational fluid dynamics, wind tunnel testing, or multiple test launches. Therefore, rather than using a control method that continuously adjusts the amount the airbrakes are deployed into the airstream, we limit our control methodology to determine a single deployment altitude with full deployment, after which the airbrakes remain deployed. If the rocket begins to undershoot the target altitude, the airbrakes retract early and remain retracted. This research investigates how various supervised machine-learning (ML) models can select the airbrake deployment altitude so that a simulated rocket reaches a 10,000-foot target apogee within one percent error. This simulation-based, quantitative, single-case design uses RocketPy, an aerospace simulation tool, to generate a large set of simulated flights with different airbrake deployment altitudes under varying environmental and launch conditions. This dataset is used to train random forest, gradient boosting, and support vector regression models to predict the optimal deployment altitude. Each model is evaluated using new RocketPy simulations with different flight conditions to compare how accurately each model achieves the target apogee. Our research question is: “Which ML model most effectively minimizes apogee error relative to the 10,000-foot target for the Calisto rocket when trained and tested through RocketPy simulations?” Model success is defined as error below one percent. This study supports future work developing more advanced ML-based controllers for physical rocket systems under site-specific weather and launch variations.
