Investigation of Mission-Driven Inverse Aircraft Design Space Exploration with Machine Learning

Abstract

The goal of this work was to investigate the feasibility of developing machine learning models for predicting the values of aircraft configuration design variables when provided with time series of mission-informed performance parameters. Regression artificial neural networks, along with their associated training data, have been generated and tested for aircraft design space exploration scenarios. The bounds of the data used to train the models were partially informed by the configuration characteristics of the Boeing 737 Next Generation family. The effects of varying neural network architecture, along with the application of different data filtering schemes, on the models’ predictive accuracy have been examined. The results obtained demonstrated that cascade-forward shallow neural networks not only exhibited excellent generalization across the design space for which the model was calibrated for, but also showcased versatility when tasked with predicting design variable values for a configuration layout relatively different from the ones used for training. Furthermore, these models had favorable metrics in computational wall-clock time required and number of epochs needed for training.

Mission-Driven Inverse Design of Blended Wing Body Aircraft with Machine Learning

“…Compared to this architecture, a cascade-forwards neural network has an additional mapping from each neuron in the input layer directly to every neuron in the output layer. This type of ANN has provided favorable predictive accuracy in scenarios involving time-series data [9]. Figure 10 illustrates the architectures for both shallow feed-forward and shallow cascade-forward ANNs.…”

“…A similar scaling operation was performed to obtain P S . Here, each time step value of every performance parameter was divided by 1.05 multiplied by the maximum value of that performance parameter within P. For example, in order to scale L/D, each time step value of L/D in every row was divided by 1.05 multiplied by the maximum value of L/D that appears in P. This operation was repeated for all 8 performance parameters in each row of P to construct P S and is expressed in Equation (9). At this point, the databases for neural network training and testing have now been generated.…”

“…The author's previously published research [9] focused on developing a framework to support the aforementioned objective, albeit within the context of applying it to conventional, TAW, single-aisle, twin-engine, commercial transport category airplane configurations such as those in the same class as the Boeing 737 Next Generation family. The work outlined in this paper is an expansion and adaptation of the authors' previous work, focusing specifically on application to the inverse design of unconventional airplane configurations, namely BWBs.…”