Application and Assessment of Machine Learning Methods for Aircraft Design Problems

In this work, we present a comparative study on machine-learning-based surrogate modeling for multidisciplinary aerospace applications. The main objective of the work is to demonstrate the predictive capability of traditional machine learning methods, including linear regression, random forest, gradient boosting, extreme gradient boosting, support vector regression, Gaussian process regression, and multilayer perceptron, on multidisciplinary engineering problems. The selected datasets are determined at different complexity levels and disciplines. These methods are first applied to two benchmark functions with different mathematical characteristics and then to four engineering design problems. The models are evaluated using different training sample sizes and multiple random data partitions to assess their predictive performance and robustness. The models are compared in terms of the coefficient of determination ([Formula: see text]), normalized root mean squared error, normalized mean absolute error, and computational cost. The obtained results show that the relative performance of the surrogate models varies with problem dimensionality and training sample size. In particular, Gaussian process regression and support vector regression provide consistently high predictive accuracy across the investigated engineering problems, while the computational requirements and sensitivity to data partitioning differ among the methods.