Estimating Underresolution Effects in Turbulence Modeling via a Data-Driven Convolutional Neural Network

Abstract

The objective of this research is to estimate and predict the effects on a computational fluid flow that gets its grid resolution downsampled. For that, direct numerical simulation (DNS) data is generated through an in-house solver for the study case of a lid-driven cavity flow. From this data in two different resolutions— one regular resolution and one downsampled by a factor of two—, a comparison is drawn in terms of how good it adapts to the downsampled DNS results between large-eddy simulation (LES) filters and a convolutional neural network (CNN) trained on this data. The results of the study give some insights about the potential use of neural networks to act as filters, where the runtime compared to DNS solver data is three orders of magnitude faster; whereas it only outperforms LES filters in 21.36% of the cases.

En este trabajo fin de grado de investigaci´on se realiza una estimaci´on de los efectos de subresoluci´on en flujos a trav´es de un enfoque basado en datos, empleando datos de un solucionador de Simulaci´on Num´erica Directa (DNS) e intentando alcanzar los resultados de menor resoluci´on a partir de los datos de mayor resoluci´on. Esta aplicaci´on tiene como objetivo reducir el coste computacional, intentando superar tanto a los c´odigos de Simulaci´on Num´erica Directa en tiempo de ejecuci´on y a los c´odigos de Simulaci´on de Grandes Torbellinos (LES) en precisi´on. Pese a que los resultados muestran que la red neuronal es tres ´ordenes de magnitud m´as r´apida que c´odigos DNS, el LES supera a la red neuronal convolucional (CNN) usada en este papel en el 78.64% de los casos.