Periodic Reporting for period 1 - DataProMat (A Diffusion Maps workflow enabled by Neural Networks and Equation-free calculations for multi scale material and process modelling)
Reporting period: 2021-09-01 to 2023-08-31
The goal is project is to develop an integrated machine learning framework for efficient material and process modeling that will help accelerate the design of new materials and their large-scale production process, by leveraging data from simulations and experiments.
Specifically:
1. To achieve Reduction of size and complexity of reactive flows through model order reduction of the reactor dynamics
2. To develop, implement, and validate a machine learning framework for the prediction of material properties.
A key component of the proposed approach is efficient dimensionality reduction using manifold learning, here by implementing Diffusion Maps. The latter can be thought of as the nonlinear counterpart of Proper Orthogonal Decomposition appropriate for data belonging to a curved manifold. This method yields a parsimonious parametrisation of data manifold, leading to improved accuracy and efficiency, in comparison to POD. Through appropriate interpolation technology, also based on Diffusion Maps, the so-called Geometric Harmonics, it is possible to map from the input to the output space (here the distribution of mass, momentum and temperature inside a chemical vapour deposition reactor) and vice versa, as well as from partial observations to other partial observations or even to the full solution space.
Specifically:
1. To achieve Reduction of size and complexity of reactive flows through model order reduction of the reactor dynamics
2. To develop, implement, and validate a machine learning framework for the prediction of material properties.
A key component of the proposed approach is efficient dimensionality reduction using manifold learning, here by implementing Diffusion Maps. The latter can be thought of as the nonlinear counterpart of Proper Orthogonal Decomposition appropriate for data belonging to a curved manifold. This method yields a parsimonious parametrisation of data manifold, leading to improved accuracy and efficiency, in comparison to POD. Through appropriate interpolation technology, also based on Diffusion Maps, the so-called Geometric Harmonics, it is possible to map from the input to the output space (here the distribution of mass, momentum and temperature inside a chemical vapour deposition reactor) and vice versa, as well as from partial observations to other partial observations or even to the full solution space.
The work is organized in 4 Work Packages:
1.2.1 Work Package 1
An appropriately reduced surrogate model of a complex deposition process, involving competing physical phenomena and chemical reactions in complex geometries, was developed. An efficient model of the chemical pathways leading to deposition on the surface was developed and validated against experimental measurements. The model accounted for only two reactions instead of the approximately 100 known reactions taking place in the deposition of alumina. An appropriately simplified representation of the real geometry is proposed, which greatly reduces the computational cost by addressing only a part of the reactor at a time. Therefore, by implementing a “divide-and-conquer” strategy it is possible to tackle the large-scale process accurately and efficiently. The accuracy of the proposed reduced models is validated through comparison to experimental measurements.
1.2.2 Work package 2
The microstructure of the deposited material and its properties was the focus of WP2. More importantly, though, the effect of various process inputs (temperature, pressure and composition of chemical compounds used as precursors) on the micro-scale properties of the material was investigated through the implementation of data-driven methods. Specifically, machine learning methods such as the Extreme Gradient Boosting (XGBOOST) algorithm as well as Artificial Neural Networks (ANNs) were trained to predict the material properties, given process inputs. A key ingredient of the proposed approach is efficient and accurate reduction of the high dimensional input and output space: nonlinear manifold learning, pursued through Diffusion Maps, can identify a low dimensional parametrization of the manifold of the data, leading to only a handful of coordinates (typically, in every case studied here, less than 10 latent variables were required for accurate prediction). This contributed to the efficient training of the machine learning models.
1.2.3 Work package 0, 3: Management, Tok, dissemination and communication
The work carried out during this project was disseminated through 5 conference and workshop talk and 4 published papers, see1.3 for details.
The two-way transfer of knowledge between the PI and the host has been summarized in Section 1.1.
The management of the project and quality assurance of the work was achieved through regular (monthly) progress meetings with the supervisor. The PI and supervisor jointly developed a career development plan for the PI.
Administrative tasks, financial management, and overall project management were done in collaboration with the University of Luxembourg. This ensured that the PI could successfully manage the execution of the project and focus his primary work on the research tasks
• Mrs. Elena Varroy, Project Coordinator supported the PI in overall project management.
• Mrs. Odile Marois, Administrative support, helped with equipment and material procurement, and organization of travel to conferences.
• Several members of IT support desk helped with equipment procurement, and support for IT related issues.
• The supervisor, Prof. Bordas, helped the PI integrate into the University of Luxembourg.
1.2.1 Work Package 1
An appropriately reduced surrogate model of a complex deposition process, involving competing physical phenomena and chemical reactions in complex geometries, was developed. An efficient model of the chemical pathways leading to deposition on the surface was developed and validated against experimental measurements. The model accounted for only two reactions instead of the approximately 100 known reactions taking place in the deposition of alumina. An appropriately simplified representation of the real geometry is proposed, which greatly reduces the computational cost by addressing only a part of the reactor at a time. Therefore, by implementing a “divide-and-conquer” strategy it is possible to tackle the large-scale process accurately and efficiently. The accuracy of the proposed reduced models is validated through comparison to experimental measurements.
1.2.2 Work package 2
The microstructure of the deposited material and its properties was the focus of WP2. More importantly, though, the effect of various process inputs (temperature, pressure and composition of chemical compounds used as precursors) on the micro-scale properties of the material was investigated through the implementation of data-driven methods. Specifically, machine learning methods such as the Extreme Gradient Boosting (XGBOOST) algorithm as well as Artificial Neural Networks (ANNs) were trained to predict the material properties, given process inputs. A key ingredient of the proposed approach is efficient and accurate reduction of the high dimensional input and output space: nonlinear manifold learning, pursued through Diffusion Maps, can identify a low dimensional parametrization of the manifold of the data, leading to only a handful of coordinates (typically, in every case studied here, less than 10 latent variables were required for accurate prediction). This contributed to the efficient training of the machine learning models.
1.2.3 Work package 0, 3: Management, Tok, dissemination and communication
The work carried out during this project was disseminated through 5 conference and workshop talk and 4 published papers, see1.3 for details.
The two-way transfer of knowledge between the PI and the host has been summarized in Section 1.1.
The management of the project and quality assurance of the work was achieved through regular (monthly) progress meetings with the supervisor. The PI and supervisor jointly developed a career development plan for the PI.
Administrative tasks, financial management, and overall project management were done in collaboration with the University of Luxembourg. This ensured that the PI could successfully manage the execution of the project and focus his primary work on the research tasks
• Mrs. Elena Varroy, Project Coordinator supported the PI in overall project management.
• Mrs. Odile Marois, Administrative support, helped with equipment and material procurement, and organization of travel to conferences.
• Several members of IT support desk helped with equipment procurement, and support for IT related issues.
• The supervisor, Prof. Bordas, helped the PI integrate into the University of Luxembourg.
The project to led to significant contributions in the field of data-enabled computational tools for large-scale processes that were published in scientific journals and presented in international conferences:
P1. Eleni D. Koronaki, Nikolaos Evangelou, Yorgos M. Psarellis, Andreas G. Boudouvis, Ioannis G. Kevrekidis, From partial data to out-of-sample parameter and observation estimation with diffusion maps and geometric harmonics, Computers & Chemical Engineering, Volume 178, 2023, 108357, ISSN 0098-1354, https://doi.org/10.1016/j.compchemeng.2023.108357.
P2. Papavasileiou, P., Koronaki, E. D., Pozzetti, G., Kathrein, M., Czettl, C., Boudouvis, A. G., & Bordas, S. P. (2023). Equation-based and data-driven modeling strategies for industrial coating processes. Computers in Industry, 149, 103938.
P3. Martin-Linares, C. P., Psarellis, Y. M., Karapetsas, G., Koronaki, E. D., & Kevrekidis, I. G. (2023). Physics-agnostic and Physics-infused machine learning for thin films flows: modeling, and predictions from small data. arXiv preprint arXiv:2301.12508.
P4. Papavasileiou, P., Koronaki, E. D., Pozzetti, G., Kathrein, M., Czettl, C., Boudouvis, A. G., & Bordas, S. P. (2022). An efficient chemistry-enhanced CFD model for the investigation of the rate-limiting mechanisms in industrial Chemical Vapor Deposition reactors. Chemical Engineering Research and Design, 186, 314-325.
C1 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, The challenges in the implementation of CFD and ML models in Industrial engineering flows, EFMC 14 – 14th European Fluid Mechanics Conference, Athens, Greece, 13- 16 September 2022.
C2 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, An efficient CFD model of an industrial scale CVD reactor allowing accurate coating thickness predictions, WCCM-APCOM 15th World Congress on Computational Mechanics and 8th Asian Pacific Congress on Computational mechanics, Yokohama 2022, 31 June-05 August 2022.
C3 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, A comparison of equation-based and machine learning models of industrial scale deposition processes, ECCOMAS2022 international congress on computational mechanics, Oslo, Norway, 5-9 June 2022.
C4 Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, S. P. A. Bordas, A. G. Boudouvis, Assessment of CFD and ML modelling strategies for industrial chemical vapor deposition reactors, 13th PESXM (Panhellenic Scientific Conference on Chemical Engineering), Patras, Greece, 2-4 June 2022.
C5 C. Martin Linares, E. D. Koronaki, Y. Psarellis, G. Karapetsas, I. G. Kevrekidis “From Navier- Stokes simulajons for thin films to amplitude equajons and back via physics-assisted machine-learning”, Bullejn of the American Physical Society, 2022.
P1. Eleni D. Koronaki, Nikolaos Evangelou, Yorgos M. Psarellis, Andreas G. Boudouvis, Ioannis G. Kevrekidis, From partial data to out-of-sample parameter and observation estimation with diffusion maps and geometric harmonics, Computers & Chemical Engineering, Volume 178, 2023, 108357, ISSN 0098-1354, https://doi.org/10.1016/j.compchemeng.2023.108357.
P2. Papavasileiou, P., Koronaki, E. D., Pozzetti, G., Kathrein, M., Czettl, C., Boudouvis, A. G., & Bordas, S. P. (2023). Equation-based and data-driven modeling strategies for industrial coating processes. Computers in Industry, 149, 103938.
P3. Martin-Linares, C. P., Psarellis, Y. M., Karapetsas, G., Koronaki, E. D., & Kevrekidis, I. G. (2023). Physics-agnostic and Physics-infused machine learning for thin films flows: modeling, and predictions from small data. arXiv preprint arXiv:2301.12508.
P4. Papavasileiou, P., Koronaki, E. D., Pozzetti, G., Kathrein, M., Czettl, C., Boudouvis, A. G., & Bordas, S. P. (2022). An efficient chemistry-enhanced CFD model for the investigation of the rate-limiting mechanisms in industrial Chemical Vapor Deposition reactors. Chemical Engineering Research and Design, 186, 314-325.
C1 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, The challenges in the implementation of CFD and ML models in Industrial engineering flows, EFMC 14 – 14th European Fluid Mechanics Conference, Athens, Greece, 13- 16 September 2022.
C2 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, An efficient CFD model of an industrial scale CVD reactor allowing accurate coating thickness predictions, WCCM-APCOM 15th World Congress on Computational Mechanics and 8th Asian Pacific Congress on Computational mechanics, Yokohama 2022, 31 June-05 August 2022.
C3 P. Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, A. G. Boudouvis, S. P. A. Bordas, A comparison of equation-based and machine learning models of industrial scale deposition processes, ECCOMAS2022 international congress on computational mechanics, Oslo, Norway, 5-9 June 2022.
C4 Papavasileiou, E. D. Koronaki, G. Pozzetti, M. Kathrein, C. Czettl, S. P. A. Bordas, A. G. Boudouvis, Assessment of CFD and ML modelling strategies for industrial chemical vapor deposition reactors, 13th PESXM (Panhellenic Scientific Conference on Chemical Engineering), Patras, Greece, 2-4 June 2022.
C5 C. Martin Linares, E. D. Koronaki, Y. Psarellis, G. Karapetsas, I. G. Kevrekidis “From Navier- Stokes simulajons for thin films to amplitude equajons and back via physics-assisted machine-learning”, Bullejn of the American Physical Society, 2022.