Machine learning enhancement of Spalart-Allmaras Turbulence Model using Convolutional Neural Network

A machine learning strategy is introduced to model a correction for the production term of the Spalart-Allmaras turbulence model. The correction function is estimated for several test cases through inverse design. The data assimilated from the inverse design process is transformed into an auxiliary representation that captures geometric information from the mean flow variables. Intuitively, the transformed data can be thought of as snapshots of local flow field information. The transformed data is used to train a convolutional neural network that can predict the correction function for general test cases. The trained machine learning model is coupled with a numerical solver such that the corrections to the Spalart-Allmaras model can be computed at every solver iteration. The performance of the machine learning enhanced model is assessed based on the accuracy of its predictions in novel flow conditions.