Data-Driven Surrogate Modeling for the Flammability Reduction System

Abstract

The aircraft fuel tank Flammability Reduction System (FRS) is a safety enhancement system that lowers flammability of the fuel tank by reducing the oxygen concentration. The certification process for this system has a unique approach involving an expensive Monte Carlo analysis, which incorporates statistical distributions of environmental, aircraft, and fuel properties. This analysis entails conducting over 10,000 missions to ensure compliance with flammability exposure requirements. To address computational demands, surrogate modeling is a popular approach to represent high-fidelity data and models through computationally efficient, albeit lower-fidelity, approximation models. This work focuses on the development of data-driven surrogate models specifically for time-series outputs obtained from FRS Monte Carlo analysis. The surrogate models are generated using a specialized artificial neural network architecture known as Long-Short Term Memory (LSTM). By utilizing simulation inputs and outputs, the LSTM-based surrogate models are able to effectively capture the behavior of the FRS system. The results demonstrate the efficacy of LSTM in modeling the FRS, and offers substantial improvements in simulation time compared to conventional methods. Despite the increased computational efficiency, the surrogate model still provides accurate estimates tailored to the specific application. Additionally, the flexible LSTM architecture is perfectly suited for training surrogate models using large databases of flight profiles characterized by non-stationary and transient responses. This suggests potential applications in the aerospace domain for modeling operational data or flight-test data, offering valuable insights into system behavior.