A QUANTITATIVE ANALYSIS OF MACHINE LEARNING BASED REGRESSORS FOR PRESSURE RECONSTRUCTION IN PARTICLE IMAGE VELOCIMETRY APPLICATIONS

摘要

With the recent advances in Artificial Intelligence, avenues for the application of Machine Learning (ML) are increasing in many engineering problems pertaining to Fluid Mechanics. Particle Image Velocimetry is an advanced fluid measurement technique but the pressure evaluation requires the data to be further processed. In the proposed work, we focus on developing a novel computational framework using Machine Learning techniques, primarily Artificial Neural Networks (ANN) to infer the pressure fields from velocity obtained from PIV data. The framework was tested for a case of flow over a periodic hill. The data for training was generated by performing LES simulations for different hill curves and Reynolds number. Various Regression models were developed in conjunction with the Adam and Limited Memory - BFGS (L-BFGS) optimizers and suitable activation functions. A detailed study of the performance of each model was done to determine the optimum values of the hyperparameters. Further, a comparative analysis of various models and the respective optimization algorithms used was performed. It was concluded that the multi-layered perceptron model with L - BFGS optimizer showed the best performance. Overall, it is observed that the models can predict the pressure values with a high degree of accuracy as confirmed by validation with the previous literature data.