Machine Learning with Supersonic Retropropulsion Wind Tunnel Test Data

Future human Mars missions will require powered descent starting at supersonic conditions, termed supersonic retropropulsion (SRP), because parachutes alone will not provide sufficient deceleration for landing human-scale payloads. We use machine learning (ML) methods to develop novel and innovative data-analytic methods to advance the fundamental understanding of multi-nozzle plume physics, quantify uncertainties, and inform planning for enabling the deployment of SRP technology. Overall, we analyze wind-tunnel test data for a wide range of input parameters and operating conditions. From the knowledge generated from the analysis of this data, robust and efficient physics-informed ML tools can be developed to support the design of SRP vehicles.

We developed a data-driven heuristic (DDH) model of the SRP quantities of interest (i.e., forces). This model uses an additive framework where we sum three terms: body drag, SRP model, and neural network residuals. The DDH model shows promising results and is capable of efficiently processing new data. More wind tunnel testing later in 2022 will provide more data for us to process through our framework to further the understanding of SRP.