Development of a Two-Level ML Spatial-temporal Framework for Industrial Thermal Striping Applications

TL;DR

This paper introduces a two-level machine learning framework that leverages physics-based insights to efficiently predict complex thermal flow patterns in industrial applications, reducing computational costs and improving accuracy.

Contribution

The paper presents a novel physics-informed two-level ML framework that overcomes limitations of traditional surrogates in nonlinear thermal flow predictions.

Findings

Accurately captures fluctuation frequencies and amplitudes in nonlinear flows

Achieves a velocity RMSE of 0.031 across 3,844 points

Demonstrates effectiveness in industrial thermal striping simulation

Abstract

A data-driven framework for spatial-temporal prediction is proposed for reducing the computational cost of industrial thermal striping applications. The framework aims to efficiently identify the flow features and utilize them in spatiotemporal field predictions with a limited number of full-order simulations. In a parameterized system, the classical projection-based surrogates often suffer from Kolmogorov n-width limitations and have limited reducibility in highly-nonlinear systems. A two-level machine learning framework is proposed based on physics to address this issue. The selection of machine learning algorithms and information extraction is empowered by the idea that the thermal striping phenomenon is driven by large turbulent coherent flow structures. In the first level, the turbulence coherent structures are identified and collected by performing Proper Orthogonal Decomposition on local parameters. A tree-based machine-learning model is then used to down-select the reference structures based on the physical meaning of the modes. In the second level, the reference structure information is broken down into pointwise local points for training a deep structure corrector, which corrects the bias between the locally true and reference structures. Demonstration of a selected industrial application, triple jet, shows that the method can capture the fluctuation frequencies and amplitudes of the spatiotemporal fields in a highly nonlinear setting. The result shows that the prediction root-mean-squared error of velocity is 0.031 for all 3,844 spatial points.