Assessing the Potential and Limitations of LSTM Inverse Models for Hybrid Rocket Throttle Control

Hybrid rockets using a liquid oxidizer and solid polymer fuel are capable of being throttled deeply without sacrificing combustion stability. However, accurately managing thrust remains difficult because of intricate relationships between the injector input and thrust output. Utah State University’s Propulsion Research Laboratory uses a digital throttle manifold for precise oxidizer flow. Throttle control uses a calibrated low-dimensional ballistics model to design thrust profiles, solving for injector pressure. This numerical inverse algorithm, referred to in this paper as the Direct Numerical Predictor (DNP), is accurate but too slow for real-time control of hardware. This paper evaluates a compact Long Short-Term Memory (LSTM) network as an inverse control model for a 75 mm GOX/ABS motor operating within a 65–225 N thrust range. A synthetic dataset is generated using various thrust profiles linked to the calibrated ballistics model. For each target thrust curve, the DNP produces a reference injector plenum pressure sequence used both for training and as a comparison baseline. The final LSTM network converts desired thrust and other selected dynamic features into injector pressure commands. Performance was tested in simulation using the DNP or LSTM commands. On a 100-profile mixed–segment stress set, the LSTM had a mean trimmed thrust root mean square error (RMSE) of 8.5 N relative to the upper-limit thrust /level and a 95th-percentile RMSE of 6.4%. Conversely, the DNP showed a mean RMSE of 1.1%. For injector pressure, the LSTM matched the DNP with a mean RMSE of 195 kPa relative to the upper-limit injector-pressure value ��UL, and a mean absolute error below 2%. The LSTM network runs approximately 100 times faster than the numerical inverse. This suggests it can efficiently serve as a rapid feedforward inverse throttle element in hybrid rockets. In addition to detailing the control simulations, this paper discusses current limitations, potential failure modes, and a structured experimental plan that incorporates hardware-in-the-loop as well as hot-fire testing.