FLOW PARAMETER PREDICTION IN DUCTS: LEVERAGING DNS DATA WITH MACHINE LEARNING

Abstract

This research study uses machine learning models to predict flow parameters like kinematic stress, pressure, and turbulence characteristics by leveraging Direct Numerical Simulation (DNS) data. Engineering applications such as aerospace, automotive, and HVAC systems require accuracy in flow parameters. For real-time applications, DNS provides details about the fluid dynamics but it is computationally expensive and not suitable. This work utilizes the data-driven approach and the use of machine learning algorithms for the prediction of flow parameters. Firstly, we pre-process and clean the DNS datasets, and input features are extracted. The dataset is divided into training subsets as well as testing subsets. At the beginning to set the baseline prediction, we use the multiple linear regression (MLR) algorithms as our initialization model. After that Random Forest (RF), an ensemble model is evaluated due to its robust predictive capabilities and less overfitting of data. To improve the model performance, we use feature engineering and dimensionality reduction techniques. These models exhibit the potential for cost-effectiveness, timesaving, and improved efficiency of simulations in engineering design processes. This study presents the ability of machine learning algorithms to act as a bridge between practical engineering needs and computational fluid dynamics. This serves as a tool for optimizing duct-based systems and understanding fluid behavior in various applications for researchers and engineers. This work contributes to combining the data-driven approaches with computational simulations for a better understanding of fluid flow in different geometries.