Extraction of Flow Features for Predicting Pressure Distribution Using Convolutional Neural Networks

EasyChair Preprint 5534

Abstract

Recent developments in Artificial Intelligence and Machine Learning have led to new sophisticated approaches to solve complex engineering problems. This study proposes an advanced Convolutional Neural Network based data driven framework to infer load or pressure values from velocity profiles. Convolutional Neural Networks are chosen over conventional Neural Networks because CNNs require reduced training variables which leads to optimized computational time and resources required for training. The model spatially extracts features from fluid flow data using convolution filters and attempts to map them with pressure profile. Hyperparameters of the model are finely tuned to ensure optimum functioning and results. Performance of the predictive model is tested on two flow cases – Converging Diverging Channel and Parametric Bump for multiple Reynolds Numbers. An opensource validated Dataset was utilized for ensuring standardized training of the model. Overall, the framework is found to be effectively efficient and a high degree of accuracy is observed in the results.

Keyphrases: Convolutional Neural Network, fluid mechanics, machine learning

@booklet{EasyChair:5534,
  author    = {Prince Tyagi and Priyam Gupta and Raj Kumar Singh},
  title     = {Extraction of Flow Features for Predicting Pressure Distribution Using Convolutional Neural Networks},
  howpublished = {EasyChair Preprint 5534},
  year      = {EasyChair, 2021}}