nach oben

Open Access 2023 | Open Access | Buch

Machine Learning and Its Application to Reacting Flows

ML and Combustion

herausgegeben von: Nedunchezhian Swaminathan, Alessandro Parente

Verlag: Springer International Publishing

Buchreihe : Lecture Notes in Energy

Enthalten in: Springer Professional "Wirtschaft+Technik" , Springer Professional "Technik" , Springer Professional "Wirtschaft"

Inhaltsverzeichnis

Über dieses Buch

This open access book introduces and explains machine learning (ML) algorithms and techniques developed for statistical inferences on a complex process or system and their applications to simulations of chemically reacting turbulent flows.

These two fields, ML and turbulent combustion, have large body of work and knowledge on their own, and this book brings them together and explain the complexities and challenges involved in applying ML techniques to simulate and study reacting flows. This is important as to the world’s total primary energy supply (TPES), since more than 90% of this supply is through combustion technologies and the non-negligible effects of combustion on environment. Although alternative technologies based on renewable energies are coming up, their shares for the TPES is are less than 5% currently and one needs a complete paradigm shift to replace combustion sources. Whether this is practical or not is entirely a different question, and an answer to this question depends on the respondent. However, a pragmatic analysis suggests that the combustion share to TPES is likely to be more than 70% even by 2070. Hence, it will be prudent to take advantage of ML techniques to improve combustion sciences and technologies so that efficient and “greener” combustion systems that are friendlier to the environment can be designed.

The book covers the current state of the art in these two topics and outlines the challenges involved, merits and drawbacks of using ML for turbulent combustion simulations including avenues which can be explored to overcome the challenges. The required mathematical equations and backgrounds are discussed with ample references for readers to find further detail if they wish. This book is unique since there is not any book with similar coverage of topics, ranging from big data analysis and machine learning algorithm to their applications for combustion science and system design for energy generation.

Inhaltsverzeichnis

Frontmatter

PDF

Open Access

Introduction

Abstract

The annual data published by IEA is analysed to get a projection for the combustion share in total primary energy supply for the world. This projection clearly identifies that more than 60% of world total primary energy supply will come from combustion based sources even in the year of 2110 despite an aggressive shift towards renewables. Hence, improving and searching for greener combustion technologies would be beneficial for addressing global warming. Computational approaches play an important role in this search. The large eddy simulation equations are presented and discussed. Potential terms which are amenable for using machine learning algorithms are identified as a prelude to later chapters of this volume.

N. Swaminathan, A. Parente

PDF Zum Volltext

Open Access

Machine Learning Techniques in Reactive Atomistic Simulations

Abstract

This chapter describes recent advances in the use of machine learning techniques in reactive atomistic simulations. In particular, it provides an overview of techniques used in training force fields with closed form potentials, developing machine-learning-based potentials, use of machine learning in accelerating the simulation process, and analytics techniques for drawing insights from simulation results. The chapter covers basic machine learning techniques, training procedures and loss functions, issues of off-line and in-lined training, and associated numerical and algorithmic issues. The chapter highlights key outstanding challenges, promising approaches, and potential future developments. While the chapter relies on reactive atomistic simulations to motivate models and methods, these are more generally applicable to other modeling paradigms for reactive flows.

H. Aktulga, V. Ravindra, A. Grama, S. Pandit

PDF Zum Volltext

Open Access

A Novel In Situ Machine Learning Framework for Intelligent Data Capture and Event Detection

Abstract

We present a novel framework for automatically detecting spatial and temporal events of interest in situ while running high performance computing (HPC) simulations. The new framework – composed from signature, measure, and decision building blocks with well-defined semantics – is tailored for parallel and distributed computing, has bounded communication and storage requirements, is generalizable to a variety of applications, and operates in an unsupervised fashion. We demonstrate the efficacy of our framework on several cases spanning scientific domains and applications of event detection: optimized input/output (I/O) in computational fluid dynamics simulations, detecting events that can lead to irreversible climate changes in simulations of polar ice sheets, and identifying optimal space-time subregions for projection-based model reduction. Additionally, we demonstrate the scalability of our framework using a HPC combustion application on the Cori supercomputer at the National Energy Research Scientific Computing Center (NERSC).

T. M. Shead, I. K. Tezaur, W. L. Davis IV, M. L. Carlson, D. M. Dunlavy, E. J. Parish, P. J. Blonigan, J. Tencer, F. Rizzi, H. Kolla

PDF Zum Volltext

Open Access

Machine-Learning for Stress Tensor Modelling in Large Eddy Simulation

Abstract

The accurate modelling of the unresolved stress tensor is particularly important for Large Eddy Simulations (LES) of turbulent flows. This term affects the transfer of energy from the largest to the smallest scales and vice versa, thus controlling the evolution of the flow field-in reacting flows, the flow field transports scalar fields such as mass fractions and temperature both of which control the species production and destruction rates. A large number of models have been developed in past years for the stress tensor in incompressible and non-reacting flows. A common characteristic of the majority of the classical models is that simplifying assumptions are typically involved in their derivation which limits their predictive ability. At the same time, various tunable parameters appear in the relevant closures whose value depends on the flow geometry/configuration/spatial location, and which require careful regularisation. Data-driven methods for the stress tensor is an emerging alternative modelling approach which may help to circumvent the above issues, and in recent studies several such models were developed and evaluated. This chapter discusses the modelling problem, presents some of the most popular algebraic models, and reviews some recent advances on data-driven methods.

Z. M. Nikolaou, Y. Minamoto, C. Chrysostomou, L. Vervisch

PDF Zum Volltext

Open Access

Machine Learning for Combustion Chemistry

Abstract

Machine learning provides a set of new tools for the analysis, reduction and acceleration of combustion chemistry. The implementation of such tools is not new. However, with the emerging techniques of deep learning, renewed interest in implementing machine learning is fast growing. In this chapter, we illustrate applications of machine learning in understanding chemistry, learning reaction rates and reaction mechanisms and in accelerating chemistry integration.

T. Echekki, A. Farooq, M. Ihme, S. M. Sarathy

PDF Zum Volltext

Open Access

Deep Convolutional Neural Networks for Subgrid-Scale Flame Wrinkling Modeling

Abstract

Subgrid-scale flame wrinkling is a key unclosed quantity for premixed turbulent combustion models in large eddy simulations. Due to the geometrical and multi-scale nature of flame wrinkling, convolutional neural networks are good candidates for data-driven modeling of flame wrinkling. This chapter presents how a deep convolutional neural network called a U-Net is trained to predict the total flame surface density from the resolved progress variable. Supervised training is performed on a database of filtered and downsampled direct numerical simulation fields. In an a priori evaluation on a slot burner configuration, the network outperforms classical dynamic models. In closing, challenges regarding the ability of deep convolutional networks to generalize to unseen configurations and their practical deployment with fluid solvers are discussed.

V. Xing, C. J. Lapeyre

PDF Zum Volltext

Open Access

Machine Learning Strategy for Subgrid Modeling of Turbulent Combustion Using Linear Eddy Mixing Based Tabulation

Abstract

This chapter describes the use of machine learning (ML) algorithms with the linear-eddy mixing (LEM) based tabulation for modeling of subgrid turbulence-chemistry interaction. The focus will be on the use of artificial neural network (ANN), particularly, supervised deep learning (DL) techniques within the finite-rate kinetics framework. We discuss the accuracy and efficiency aspects of two different strategies, where LEM based tabulation is used in both of them. While in the first approach, referred to as LANN-LES, the subgrid reaction-rate term is obtained efficiently using ANN in the conventional LEMLES framework, in the other approach referred to as TANN-LES, the filtered reaction rate terms are obtained using ANN. First, we assess the implications of the employed network architecture, and the associated hyperparameters, such as the amount of training and test data, epoch, optimizer, learning rate, sample size, etc. Afterward, the effectiveness of the two strategies is examined by comparing with conventional LES and LEMLES approaches by considering canonical premixed and non-premixed configurations. Finally, we describe the key challenges and future outlook of the use of ML based subgrid modelling within the finite-rate kinetics framework.

R. Ranjan, A. Panchal, S. Karpe, S. Menon

PDF Zum Volltext

Open Access

On the Use of Machine Learning for Subgrid Scale Filtered Density Function Modelling in Large Eddy Simulations of Combustion Systems

Abstract

The application of machine learning algorithms to model subgrid-scale filtered density functions (FDFs), required to estimate filtered reaction rates for Large Eddy Simulation (LES) of chemically reacting flows, is discussed in this chapter. Three test cases, i.e., a low-swirl premixed methane-air flame, a MILD combustion of methane-air mixtures, and a kerosene spray turbulent flame, are presented. The scalar statistics in these test cases may not be easily represented using the commonly used presumed shapes for modeling FDFs of mixture fraction and progress variable. Hence, the use of ML methods is explored. Particularly, deep neural network (DNN) to infer joint FDFs of mixture fraction and progress variable is reviewed here. The Direct Numerical Simulation (DNS) datasets employed to train the DNNs in each test case are described. The DNN performances are shown and compared to typical presumed probability density function (PDF) models. Finally, this chapter examines the advantages and caveats of the DNN-based approach.

S. Iavarone, H. Yang, Z. Li, Z. X. Chen, N. Swaminathan

PDF Zum Volltext

Open Access

Reduced-Order Modeling of Reacting Flows Using Data-Driven Approaches

Abstract

Data-driven modeling of complex dynamical systems is becoming increasingly popular across various domains of science and engineering. This is thanks to advances in numerical computing, which provides high fidelity data, and to algorithm development in data science and machine learning. Simulations of multicomponent reacting flows can particularly profit from data-based reduced-order modeling (ROM). The original system of coupled partial differential equations that describes a reacting flow is often large due to high number of chemical species involved. While the datasets from reacting flow simulation have high state-space dimensionality, they also exhibit attracting low-dimensional manifolds (LDMs). Data-driven approaches can be used to obtain and parameterize these LDMs. Evolving the reacting system using a smaller number of parameters can yield substantial model reduction and savings in computational cost. In this chapter, we review recent advances in ROM of turbulent reacting flows. We demonstrate the entire ROM workflow with a particular focus on obtaining the training datasets and data science and machine learning techniques such as dimensionality reduction and nonlinear regression. We present recent results from ROM-based simulations of experimentally measured Sandia flames D and F. We also delineate a few remaining challenges and possible future directions to address them. This chapter is accompanied by illustrative examples using the recently developed Python software, PCAfold. The software can be used to obtain, analyze and improve low-dimensional data representations. The examples provided herein can be helpful to students and researchers learning to apply dimensionality reduction, manifold approaches and nonlinear regression to their problems. The Jupyter notebook with the examples shown in this chapter can be found on GitHub at https://github.com/kamilazdybal/ROM-of-reacting-flows-Springer.

K. Zdybał, M. R. Malik, A. Coussement, J. C. Sutherland, A. Parente

PDF Zum Volltext

Open Access

AI Super-Resolution: Application to Turbulence and Combustion

Abstract

This article summarizes and discusses recent developments with respect to artificial intelligence (AI) super-resolution as a subfilter model for large-eddy simulations. The focus is on the application of physics-informed enhanced super-resolution generative adversarial networks (PIESRGANs) for subfilter closure in turbulence and combustion applications. A priori and a posteriori results are presented for various applications, ranging from decaying turbulence to finite-rate chemistry flows. The high accuracy of AI super-resolution-based subfilter models is emphasized, and advantages and shortcoming are described.

M. Bode

PDF Zum Volltext

Open Access

Machine Learning for Thermoacoustics

Abstract

This chapter demonstrates three promising ways to combine machine learning with physics-based modelling in order to model, forecast, and avoid thermoacoustic instability. The first method assimilates experimental data into candidate physics-based models and is demonstrated on a Rijke tube. This uses Bayesian inference to select the most likely model. This turns qualitatively-accurate models into quantitatively-accurate models that can extrapolate, which can be combined powerfully with automated design. The second method assimilates experimental data into level set numerical simulations of a premixed bunsen flame and a bluff-body stabilized flame. This uses either an Ensemble Kalman filter, which requires no prior simulation but is slow, or a Bayesian Neural Network Ensemble, which is fast but requires prior simulation. This method deduces the simulations’ parameters that best reproduce the data and quantifies their uncertainties. The third method recognises precursors of thermoacoustic instability from pressure measurements. It is demonstrated on a turbulent bunsen flame, an industrial fuel spray nozzle, and full scale aeroplane engines. With this method, Bayesian Neural Network Ensembles determine how far each system is from instability. The trained BayNNEs out-perform physics-based methods on a given system. This method will be useful for practical avoidance of thermoacoustic instability.

Matthew P. Juniper

PDF Zum Volltext

Backmatter

PDF

Titel: Machine Learning and Its Application to Reacting Flows
herausgegeben von: Nedunchezhian Swaminathan
Alessandro Parente
Verlag: Springer International Publishing
Electronic ISBN: 978-3-031-16248-0
Print ISBN: 978-3-031-16247-3
DOI: https://doi.org/10.1007/978-3-031-16248-0