Application of Multi-Task Learning Architectures to Aerodynamic Surrogate Modeling

Advancements in the area of machine learning and data science have recently motivated an increased interest in the use of these technologies for application in aerodynamic modeling. Aerodynamic surrogate models seek to utilize a set of data obtained from a high fidelity physics solver to train a fast-running model that is capable of producing the same quantities of interest with a high degree of accuracy and a fraction of the computational cost. Often, these application areas require models that can predict multiple quantities of interest (QOIs) that are related to one another. Previous surrogate modeling approaches have had minimal consideration of the co-varying nature of the QOIs by applying a simple model architecture to the prediction of these quantities in an array or creating independent models for each. In this study, we apply the cross-stitch network architecture of Misra et al. to the learning of aerodynamic force coefficients. Computational experiments are performed on the PALMO benchmark dataset and a three-dimensional missile geometry dataset. It is shown that the architecture of these networks produces more accurate and efficient surrogate models as compared to traditional multilayer perceptron architectures due to the models' ability to share information common to the prediction of multiple QOIs while still retaining task-specificity in the architecture.