國立虎尾科技大學 |

Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks./
作者:	Almeldein, Ahmed M. Z.
面頁冊數:	1 online resource (118 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
Contained By:	Dissertations Abstracts International85-11B.
標題:	Chemical engineering. -
電子資源:	click for full text (PQDT)
ISBN:	9798382736143

Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks.
Almeldein, Ahmed M. Z.

Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks. - 1 online resource (118 pages)

Source: Dissertations Abstracts International, Volume: 85-11, Section: B.

Thesis (Ph.D.)--University of Massachusetts Lowell, 2024.

Includes bibliographical references

In a technological era driven by combustion, combustion simulation, particularly CFD simulation of reacting flows, has become a very effective tool in designing and optimizing everyday technologies. A detailed and accurate combustion simulation tool is essential for cheaper, more efficient, and green transportation/energy. Detailed chemical kinetics calculations can be very computationally expensive, so various approaches have been used to speed up combustion calculations. Artificial Neural Networks (ANNs) are one promising approach that has seen significant development recently. Despite the increasing effort placed in this area, there are still uncertainties about how best to construct an ANN to reproduce the highly non-linear and numerically stiff system representing chemical kinetics calculations. This study integrates two ANN concepts to accelerate the determination of chemical kinetics while maintaining the accuracy and satisfying physical governing equations.Two ANN concepts were implemented in developing the networks for chemical kinetics applications: Mixture-of-Experts (MoE) and Physics Informed Neural Network (PINN). MoE is an ANN architecture constructed as an ensemble of multiple subnetworks, i.e., experts, where experts are trained either independently or simultaneously. Then, the outputs from those experts are combined to generate the final MOE prediction. The MoE architecture allows the different sub-networks to specialize in different thermochemical regimes, such as pre-ignition, flame front, post-flame, and equilibrium chemistry, reducing the overall size of the network needed compared with a single monolithic network for all conditions. PINNs are a class of neural networks with physical governing equations embedded within the training process to create networks that follow those governing equations. A new MoE-based PINNs approach to chemical kinetics modeling has been developed to incorporate the conservation of atomic species and total energy in the reacting mixture, i.e., summation of all sensible and chemical energy. In addition to networks trained without physical governing equations, three variations on PINNs were studied: Networks implementing conservation of atomic species only, networks implementing total energy only, and networks implementing both governing equations.MoE networks of various sizes, with between 3 and 20 experts and 6,600 to 77,000 neurons, were developed with the four PINN configurations: without physics, with conservation of atomic species only, with conservation of total energy only, and with both conservation equations. Then, networks were trained on predicting the thermochemical state after constant timesteps of 1 x 10−6 s and 1 x 10−8 s and tested using 0D constant volume combustion simulations. The ANNs trained at 1 x 10−6 s were found to reproduce the accuracy of standard kinetics solvers while satisfying the conservation of atomic mass and energy equations. They were found to achieve very accurate ignition delay predictions. On the other hand, networks trained on 1 x 10−8 were found to have lower prediction accuracy and failed to predict the ignition delay accurately. However, their smaller time step makes them more accurate in the flam front or regions where the reaction rates are very high.The ANNs with timesteps of 1 x 10−6 s that passed both one-step prediction evaluation and 0D simulation were found to have a 10-50x speed-up in online evaluation time when using CPUs and a speed-up of over 200x for the average network when using a GPU. That translates to a potential reduction in the detailed chemistry simulation time of 90%-98% for CPUs and more than 99.5% for GPUs.This approach promises ANN as a much computationally cheaper replacement for chemical kinetics solvers in combustion simulation. When linked with CFD solvers, the presented MoE-based PINNs approach is to speed up the simulation process while maintaining the accuracy of detailed chemical kinetics, thus accelerating the ongoing studies of new environmentally friendly fuels and engines.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798382736143Subjects--Topical Terms:

555952
Chemical engineering.
Subjects--Index Terms:

Artificial Neural NetworksIndex Terms--Genre/Form:

554714
Electronic books.

Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks.
LDR:05503ntm a22003977 4500 001 1144988
005 20240617111735.5
006 m o d
007 cr mn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798382736143
035 $a (MiAaPQ)AAI31146969
035 $a AAI31146969
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Almeldein, Ahmed M. Z. $3 1470182
245 1 0 $a Accelerating the Chemical Kinetics Calculations in Combustion Simulations Using Physics Informed Neural Networks.
264 0 $c 2024
300 $a 1 online resource (118 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
500 $a Advisor: Van Dam, Noah E.
502 $a Thesis (Ph.D.)--University of Massachusetts Lowell, 2024.
504 $a Includes bibliographical references
520 $a In a technological era driven by combustion, combustion simulation, particularly CFD simulation of reacting flows, has become a very effective tool in designing and optimizing everyday technologies. A detailed and accurate combustion simulation tool is essential for cheaper, more efficient, and green transportation/energy. Detailed chemical kinetics calculations can be very computationally expensive, so various approaches have been used to speed up combustion calculations. Artificial Neural Networks (ANNs) are one promising approach that has seen significant development recently. Despite the increasing effort placed in this area, there are still uncertainties about how best to construct an ANN to reproduce the highly non-linear and numerically stiff system representing chemical kinetics calculations. This study integrates two ANN concepts to accelerate the determination of chemical kinetics while maintaining the accuracy and satisfying physical governing equations.Two ANN concepts were implemented in developing the networks for chemical kinetics applications: Mixture-of-Experts (MoE) and Physics Informed Neural Network (PINN). MoE is an ANN architecture constructed as an ensemble of multiple subnetworks, i.e., experts, where experts are trained either independently or simultaneously. Then, the outputs from those experts are combined to generate the final MOE prediction. The MoE architecture allows the different sub-networks to specialize in different thermochemical regimes, such as pre-ignition, flame front, post-flame, and equilibrium chemistry, reducing the overall size of the network needed compared with a single monolithic network for all conditions. PINNs are a class of neural networks with physical governing equations embedded within the training process to create networks that follow those governing equations. A new MoE-based PINNs approach to chemical kinetics modeling has been developed to incorporate the conservation of atomic species and total energy in the reacting mixture, i.e., summation of all sensible and chemical energy. In addition to networks trained without physical governing equations, three variations on PINNs were studied: Networks implementing conservation of atomic species only, networks implementing total energy only, and networks implementing both governing equations.MoE networks of various sizes, with between 3 and 20 experts and 6,600 to 77,000 neurons, were developed with the four PINN configurations: without physics, with conservation of atomic species only, with conservation of total energy only, and with both conservation equations. Then, networks were trained on predicting the thermochemical state after constant timesteps of 1 x 10−6 s and 1 x 10−8 s and tested using 0D constant volume combustion simulations. The ANNs trained at 1 x 10−6 s were found to reproduce the accuracy of standard kinetics solvers while satisfying the conservation of atomic mass and energy equations. They were found to achieve very accurate ignition delay predictions. On the other hand, networks trained on 1 x 10−8 were found to have lower prediction accuracy and failed to predict the ignition delay accurately. However, their smaller time step makes them more accurate in the flam front or regions where the reaction rates are very high.The ANNs with timesteps of 1 x 10−6 s that passed both one-step prediction evaluation and 0D simulation were found to have a 10-50x speed-up in online evaluation time when using CPUs and a speed-up of over 200x for the average network when using a GPU. That translates to a potential reduction in the detailed chemistry simulation time of 90%-98% for CPUs and more than 99.5% for GPUs.This approach promises ANN as a much computationally cheaper replacement for chemical kinetics solvers in combustion simulation. When linked with CFD solvers, the presented MoE-based PINNs approach is to speed up the simulation process while maintaining the accuracy of detailed chemical kinetics, thus accelerating the ongoing studies of new environmentally friendly fuels and engines.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Chemical engineering. $3 555952
650 4 $a Mechanical engineering. $3 557493
653 $a Artificial Neural Networks
653 $a Chemical kinetics
653 $a Combustion
653 $a Methane
653 $a Kinetics solvers
655 7 $a Electronic books. $2 local $3 554714
690 $a 0548
690 $a 0542
690 $a 0800
710 2 $a University of Massachusetts Lowell. $b Mechanical Engineering. $3 1241347
710 2 $a ProQuest Information and Learning Co. $3 1178819
773 0 $t Dissertations Abstracts International $g 85-11B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31146969 $z click for full text (PQDT)