An improved machine learning method for thermochemistry tabulation, with application to LES-PDF simulations of piloted diffusion and swirl-bluff-body stabilised flames with NOx formation

IF 5.8 2区工程技术 Q2 ENERGY & FUELS

Combustion and Flame Pub Date : 2025-04-15 DOI:10.1016/j.combustflame.2025.114130

Tianjie Ding, W.P. Jones, Stelios Rigopoulos

{"title":"An improved machine learning method for thermochemistry tabulation, with application to LES-PDF simulations of piloted diffusion and swirl-bluff-body stabilised flames with NOx formation","authors":"Tianjie Ding, W.P. Jones, Stelios Rigopoulos","doi":"10.1016/j.combustflame.2025.114130","DOIUrl":null,"url":null,"abstract":"<div><div>Many turbulent combustion modelling approaches require real-time computation of reaction source terms, which is time-consuming and represents a bottleneck of turbulent combustion simulations. In order to speed up the reaction computation process, an artificial neural network (ANN) thermochemistry tabulation methodology has been proposed and developed in our previous work (Ding et al., 2021). In the present work, this methodology is further developed and applied to thermochemistry that includes NO<sub>x</sub> formation, which poses further challenges due to the small concentrations of the N-containing species. In particular, we build on the Multiple Multilayer Perceptrons (MMLP) concept, which aims to improve prediction accuracy by systematically combining multiple ANNs. A new method, MMLP-II, is proposed in this work, which trains different ANNs to predict states with different ranges of initial species concentration, in contrast to the previous MMLP-I method which trains several ANNs with different ranges of output magnitude. Both MMLP methods are applied to tabulate the complete GRI-3.0 mechanism and the resulting ANNs are tested on two different turbulent methane flames: Sandia flame D and Sydney flame SMA2. It is found that MMLP-II method can reduce the ANN error accumulation of minor species, and very accurate results are obtained in both turbulent flames. The successful application to two different turbulent combustion problems is indicative of the capacity for generalisation of the ANN tabulation approach. Finally, the reaction integration step is accelerated by a factor of about 15 with ANNs, thus rendering chemical kinetics no longer the bottleneck of the whole simulation.</div><div><strong>Novelty and significance</strong></div><div>A new ANN thermochemistry tabulation methodology is developed, aimed at providing the higher accuracy needed for predicting species with small concentrations, as encountered in NO<sub>x</sub> chemistry. The methodology is applied to two different turbulent flames with NO<sub>x</sub> formation: a piloted diffusion flame (Sandia D) and a swirl-bluff-body stabilised flame (Sydney SMA2). The results demonstrate that the method provides both high accuracy and capacity for generalisation. The LES-PDF method is employed here, but the method is applicable to any method that involves real-time calculation of thermochemistry.</div></div>","PeriodicalId":280,"journal":{"name":"Combustion and Flame","volume":"277 ","pages":"Article 114130"},"PeriodicalIF":5.8000,"publicationDate":"2025-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Combustion and Flame","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0010218025001683","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENERGY & FUELS","Score":null,"Total":0}

引用次数: 0

Abstract

Many turbulent combustion modelling approaches require real-time computation of reaction source terms, which is time-consuming and represents a bottleneck of turbulent combustion simulations. In order to speed up the reaction computation process, an artificial neural network (ANN) thermochemistry tabulation methodology has been proposed and developed in our previous work (Ding et al., 2021). In the present work, this methodology is further developed and applied to thermochemistry that includes NO_x formation, which poses further challenges due to the small concentrations of the N-containing species. In particular, we build on the Multiple Multilayer Perceptrons (MMLP) concept, which aims to improve prediction accuracy by systematically combining multiple ANNs. A new method, MMLP-II, is proposed in this work, which trains different ANNs to predict states with different ranges of initial species concentration, in contrast to the previous MMLP-I method which trains several ANNs with different ranges of output magnitude. Both MMLP methods are applied to tabulate the complete GRI-3.0 mechanism and the resulting ANNs are tested on two different turbulent methane flames: Sandia flame D and Sydney flame SMA2. It is found that MMLP-II method can reduce the ANN error accumulation of minor species, and very accurate results are obtained in both turbulent flames. The successful application to two different turbulent combustion problems is indicative of the capacity for generalisation of the ANN tabulation approach. Finally, the reaction integration step is accelerated by a factor of about 15 with ANNs, thus rendering chemical kinetics no longer the bottleneck of the whole simulation.

Novelty and significance

A new ANN thermochemistry tabulation methodology is developed, aimed at providing the higher accuracy needed for predicting species with small concentrations, as encountered in NO_x chemistry. The methodology is applied to two different turbulent flames with NO_x formation: a piloted diffusion flame (Sandia D) and a swirl-bluff-body stabilised flame (Sydney SMA2). The results demonstrate that the method provides both high accuracy and capacity for generalisation. The LES-PDF method is employed here, but the method is applicable to any method that involves real-time calculation of thermochemistry.

查看原文本刊更多论文

一种用于热化学制表的改进机器学习方法，并应用于具有NOx形成的导航扩散和漩涡-崖体稳定火焰的LES-PDF模拟

许多湍流燃烧建模方法需要实时计算反应源项，这非常耗时，是湍流燃烧模拟的瓶颈。为了加快反应计算过程，我们在之前的工作（Ding 等人，2021 年）中提出并开发了一种人工神经网络（ANN）热化学制表方法。在本研究中，我们进一步开发了这一方法，并将其应用于包括氮氧化物形成在内的热化学过程，由于含 N 物种的浓度较小，这给研究带来了更多挑战。特别是，我们以多层感知器（MMLP）概念为基础，旨在通过系统地组合多个 ANN 来提高预测精度。本文提出了一种新方法 MMLP-II，与之前的 MMLP-I 方法不同的是，MMLP-II 会训练不同的 ANN，以预测不同初始物种浓度范围的状态，而 MMLP-I 方法则会训练不同输出幅度范围的多个 ANN。这两种 MMLP 方法都适用于表列完整的 GRI-3.0 机制，并在两种不同的湍流甲烷火焰上测试了由此产生的 ANN：桑迪亚火焰 D 和悉尼火焰 SMA2。结果发现，MMLP-II 方法可以减少 ANN 误差积累的次要物种，并且在两种湍流火焰中都获得了非常精确的结果。对两种不同湍流燃烧问题的成功应用表明，ANN 表格法具有通用性。最后，ANN 的反应整合步骤加快了约 15 倍，从而使化学动力学不再是整个模拟的瓶颈。新颖性和意义开发了一种新的 ANN 热化学制表方法，旨在为预测氮氧化物化学中遇到的小浓度物种提供所需的更高精度。该方法适用于两种不同的氮氧化物形成湍流火焰：先导扩散火焰（桑迪亚 D）和漩涡-湍流-稳定体火焰（悉尼 SMA2）。结果表明，该方法具有高精度和通用性。这里采用的是 LES-PDF 方法，但该方法适用于任何涉及热化学实时计算的方法。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Combustion and Flame 工程技术-工程：化工

CiteScore

9.50

自引率

20.50%

发文量

631

审稿时长

3.8 months

期刊介绍： The mission of the journal is to publish high quality work from experimental, theoretical, and computational investigations on the fundamentals of combustion phenomena and closely allied matters. While submissions in all pertinent areas are welcomed, past and recent focus of the journal has been on: Development and validation of reaction kinetics, reduction of reaction mechanisms and modeling of combustion systems, including: Conventional, alternative and surrogate fuels; Pollutants; Particulate and aerosol formation and abatement; Heterogeneous processes. Experimental, theoretical, and computational studies of laminar and turbulent combustion phenomena, including: Premixed and non-premixed flames; Ignition and extinction phenomena; Flame propagation; Flame structure; Instabilities and swirl; Flame spread; Multi-phase reactants. Advances in diagnostic and computational methods in combustion, including: Measurement and simulation of scalar and vector properties; Novel techniques; State-of-the art applications. Fundamental investigations of combustion technologies and systems, including: Internal combustion engines; Gas turbines; Small- and large-scale stationary combustion and power generation; Catalytic combustion; Combustion synthesis; Combustion under extreme conditions; New concepts.