Physics-integrated machine learning: embedding a neural network in the Navier-Stokes equations. Part II

TL;DR

This paper advances physics-integrated machine learning by embedding neural networks within Navier-Stokes equations using TensorFlow, enabling training without labeled data and demonstrating potential for turbulence modeling.

Contribution

It introduces a framework for training neural networks embedded in PDEs without labeled outputs, leveraging PDE solutions to guide learning.

Findings

Neural networks can be trained directly on PDE solutions without labeled data.

The framework successfully predicts turbulent viscosity and Reynolds stress derivatives.

Potential for developing PDE-based closure models for turbulence.

Abstract

The work is a continuation of a paper by Iskhakov A.S. and Dinh N.T. "Physics-integrated machine learning: embedding a neural network in the Navier-Stokes equations". Part I // arXiv:2008.10509 (2020) [1]. The proposed in [1] physics-integrated (or PDE-integrated (partial differential equation)) machine learning (ML) framework is furtherly investigated. The Navier-Stokes equations are solved using the Tensorflow ML library for Python programming language via the Chorin's projection method. The Tensorflow solution is integrated with a deep feedforward neural network (DFNN). Such integration allows one to train a DFNN embedded in the Navier-Stokes equations without having the target (labeled training) data for the direct outputs from the DFNN; instead, the DFNN is trained on the field variables (quantities of interest), which are solutions for the Navier-Stokes equations (velocity and pressure fields). To demonstrate performance of the framework, two additional case studies are formulated: 2D turbulent lid-driven cavities with predicted by a DFNN (a) turbulent viscosity and (b) derivatives of the Reynolds stresses. Despite its complexity and computational cost, the proposed physics-integrated ML shows a potential to develop a "PDE-integrated" closure relations for turbulent models and offers principal advantages, namely: (i) the target outputs (labeled training data) for a DFNN might be unknown and can be recovered using the knowledge base (PDEs); (ii) it is not necessary to extract and preprocess information (training targets) from big data, instead it can be extracted by PDEs; (iii) there is no need to employ a physics- or scale-separation assumptions to build a closure model for PDEs. The advantage (i) is demonstrated in the Part I paper [1], while the advantage (ii) is the subject of the current paper.