DEEPMECH: Deep-learning methods to tackle outstanding problems in engineering mechanics
Title: DEEPMECH: Deep-learning methods to tackle outstanding problems in engineering mechanics
SNIC Project: Berzelius-2021-96
Project Type: LiU Berzelius
Principal Investigator: Ricardo Vinuesa <rvinuesa@mech.kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2021-12-16 – 2022-07-01
Classification: 10105
Homepage: https://www.vinuesalab.com/
Keywords:

Abstract

This project aims to tackle three different problems and solve them by developing numerical methods based ondeep neural network architectures and artificial intelligence. The first problem is to predict wall-parallel velocity fields from coarse wall measurements of shear stress and pressure. The PI and his team have developed a numerical model based on a deep neural network architecture to reconstruct turbulent-flow quantities from coarse wallmeasurements in a channel flow [1-4].The project aims to add the temporal information obtained in the aforementioned model and develop a new model for further improvements. The analysis will be performed based on a database of a turbulent open-channel flow with specific friction Reynolds numbers generated through direct numerical simulations (DNS).Coarse wall measurements will be generated with different down-sampling factors from the high-resolution fields,and wall-parallel velocity fields will be reconstructed at inner-scaled wall-normal distances. The proposed method can be used to enhance the resolution of coarse wall measurements and hasexcellent potentialfor closed-loop control applications relying on non-intrusive sensing.Although the predictions in the previous study are in good agreement with large-scale patterns obtained from the filtered DNS reference, the computational cost and requirements for producing these data increase for real-life applications with higher Reynolds numbers and predictions are almost impossible. Consequently, in this study,we are looking for alternatives by combining different neural networks to improve the methodology for solving real-life scenarios. In the second problem, we are trying to predict the flow pattern in urban environmentsdue to its significant impact on air quality and thermal effects in cities worldwide. The PI and his team have provided an overview of efforts based on experiments and simulations to achieveperception into thiscomplex physical phenomenon[5]. They also investigatedthe mean flow and turbulence statistics of the flow through a simplified urban environment [6]. These studies are helpful for civil engineering, pedestrian comfort, and health concerns caused by pollutant spreading. They also suggest apervasive use of data-driven methods to characterize flow structures to furtherunderstand the dynamics of urban flows, intending to confront the crucial sustainabilitychallenges associated with them. We believe that artificial intelligence and urban flows should be combined and start a new research line, where classical data-driven tools and machine-learning algorithms can enlighten the physical mechanisms associated with urban pollution. This study is going to develop a model to predict the urban flow through a simple configuration of squared blocks, representing the buildings in a city. This model will utilize the temporal information obtained by the model presented in the previous studies [5,6] and aims to improve the predictions by investigating different neural network architectures. As a result, more complex and realistic scenarios providing specific ideas of the pollution dispersion in real cities can be presented. Although the existing approaches to predict urban flow have been barely explored due to the high complexity of the problem, using new data-driventechniques will shed light into flow physics and will provide general descriptions ofthe main mechanisms involved in the flow dynamics in urban flows.