Periodic Reporting for period 1 - REDAFLOW (REconstruction-based DAta-assisted frameworks for turbulent reacting FLOWs.)
Reporting period: 2021-09-01 to 2024-08-31
The fellow Dr. Zacharias M. Nikolaou conducted a project to develop a generalised and computationally efficient modelling framework for simulating turbulent and reacting flows by employing two promising methods: signal reconstruction/deconvolution and machine learning.
The work was carried out at the CORIA lab, a joint research unit between CNRS, the University of Rouen, and INSA Rouen, under the supervision of Prof. Luc Vervisch.
The proposed modelling frameworks had to enable complex reacting flow simulations to be conducted under a unified modelling framework.
This had to allow robust and accurate simulations of challenging flows in industrial engineering devices, thereby improving the modelling capabilities for the design of the next generation of greener energy technologies such as gas-turbine combustors, reformers, etc.
The work was carried out at the CORIA lab, a joint research unit between CNRS, the University of Rouen, and INSA Rouen, under the supervision of Prof. Luc Vervisch.
The proposed modelling frameworks had to enable complex reacting flow simulations to be conducted under a unified modelling framework.
This had to allow robust and accurate simulations of challenging flows in industrial engineering devices, thereby improving the modelling capabilities for the design of the next generation of greener energy technologies such as gas-turbine combustors, reformers, etc.
The fellow has produced three seminal publications as detailed in the publications section.
1. Combust. Flame 246 (2022) 112425. Criteria to switch from tabulation to neural networks in computational combustion
In this seminal work, the fellow has addressed a long-standing issue in the use of machine-learning over the use
of tabulation methods for solving generalised regression problems. In this work, theoretical bounds were developed
which quantify the performance of neural-networks over tabulation methods. The simple analytical formulas
essentially dictate which of the two methods is best to solve a regression problem. It was shown for instance
that neural-networks offer significant time/memory advantages for relatively large-parameter problems whereas
for small-parameter problems tabulation methods are more suitable to use. This work was also presented at the
19th international conference of numerical combustion (ICNC2024).
2. Comp. Fluids 255 (2023) 105840. An optimisation framework for the development of explicit discrete forward and inverse filters.
In this work, robust and efficient direct-inversion discrete filters were developed along with a reconstruction library.
The new filters allow filtered signals to be reconstructed at ease with order-of-magnitude time savings in comparison to
classic reconstruction methods. The new filters were also used to develop generalised reconstruction-based modelling
frameworks, and were tested using high-order direct simulation data.
3. Journal Fluid Mech. 983 (2024) A47. Revisiting the modelling framework for the unresolved scalar variance.
In this work, a novel theoretical modelling framework based on the concept of reconstruction was developed.
This new modelling framework is generalised, makes no assumptions about the underlying flow field, and
can be used to develop simulation models of arbitrary accuracy .Specifically, a variety of low-order and high-order
models for the unresolved scalar variance were developed as well as new models based on reconstruction.
The new models outperformed all classic models in the literature as evidenced by thorough testing using
high-order direct numerical simulation data.
1. Combust. Flame 246 (2022) 112425. Criteria to switch from tabulation to neural networks in computational combustion
In this seminal work, the fellow has addressed a long-standing issue in the use of machine-learning over the use
of tabulation methods for solving generalised regression problems. In this work, theoretical bounds were developed
which quantify the performance of neural-networks over tabulation methods. The simple analytical formulas
essentially dictate which of the two methods is best to solve a regression problem. It was shown for instance
that neural-networks offer significant time/memory advantages for relatively large-parameter problems whereas
for small-parameter problems tabulation methods are more suitable to use. This work was also presented at the
19th international conference of numerical combustion (ICNC2024).
2. Comp. Fluids 255 (2023) 105840. An optimisation framework for the development of explicit discrete forward and inverse filters.
In this work, robust and efficient direct-inversion discrete filters were developed along with a reconstruction library.
The new filters allow filtered signals to be reconstructed at ease with order-of-magnitude time savings in comparison to
classic reconstruction methods. The new filters were also used to develop generalised reconstruction-based modelling
frameworks, and were tested using high-order direct simulation data.
3. Journal Fluid Mech. 983 (2024) A47. Revisiting the modelling framework for the unresolved scalar variance.
In this work, a novel theoretical modelling framework based on the concept of reconstruction was developed.
This new modelling framework is generalised, makes no assumptions about the underlying flow field, and
can be used to develop simulation models of arbitrary accuracy .Specifically, a variety of low-order and high-order
models for the unresolved scalar variance were developed as well as new models based on reconstruction.
The new models outperformed all classic models in the literature as evidenced by thorough testing using
high-order direct numerical simulation data.
The three seminal publications produced thus far have reset the reacting flow modelling field. The standard practice in the literature thus far was to employ numerous models for each term in the governing equations
for large-eddy simulations. These classic models were based on limiting assumptions and included numerous parameters which required tuning in order to produce useful results. In the REDAFLOW project
a new paradigm for model development was developed without making any explicit assumptions. Under this new modelling paradigm virtually any term can be modelling under a unified and robust
mathematical framework. At the same time, the project has resulted in a new approach for assessing the computational efficiency of neural-network-based models which takes into account the structure
of the network, the number of inputs, and the number of outputs. The theoretical bounds derived on the neural-network performance are seminal, and have not been addressed before in the literature.
Going forward, we expect the newly-developed models to be adopted by both industry and academia, and relevant libraries have already been developed.
for large-eddy simulations. These classic models were based on limiting assumptions and included numerous parameters which required tuning in order to produce useful results. In the REDAFLOW project
a new paradigm for model development was developed without making any explicit assumptions. Under this new modelling paradigm virtually any term can be modelling under a unified and robust
mathematical framework. At the same time, the project has resulted in a new approach for assessing the computational efficiency of neural-network-based models which takes into account the structure
of the network, the number of inputs, and the number of outputs. The theoretical bounds derived on the neural-network performance are seminal, and have not been addressed before in the literature.
Going forward, we expect the newly-developed models to be adopted by both industry and academia, and relevant libraries have already been developed.