Periodic Reporting for period 1 - ConvExt (EXTreme events in CONVection: advanced measurements and data-driven prediction)
Período documentado: 2021-06-01 hasta 2023-05-31
Due to the need of an environmentally sustainable energy production, the generation of electricity from wind power is expanding worldwide, with Europe being the second for installed capacity after Asia. The optimal placement of wind turbines requires the knowledge of wind conditions at any place of a terrain down to millimeter scales. Large wind fluctuations at small scales, also called wind gusts are extreme events responsible for the phenomenon of intermittency in atmospheric turbulence. Intermittency is not considered by current flow models of atmospheric turbulence, even if it may have important consequences for wind turbines operating conditions. For example, it was observed that the electrical power fed into the grid by a wind farm may change by 50% within two minutes because of wind gusts.
This project aims at understanding and predicting extreme events of energy dissipation from experiments in Rayleigh-Bénard convection (RBC), a turbulent fluid layer that is uniformly heated from below and cooled from above. RBC is considered as a paradigm for atmospheric turbulence. Measurements of extreme events of energy dissipation are performed using Particle Image Velocimetry (PIV).
Machine Learning (ML) methods are an efficient and unbiased way to process a large amount of data effectively. ML has been rarely applied to the detection of extreme events in fluid turbulence. In this project, by using Recurrent Neural Networks (a particular king of ML method), a large amount of data from experiments and simulations is processed to construct a model. The predicted extreme events are large wind fluctuations at small scales, wind gusts, which may have important consequences for wind turbines operating conditions.
In summary, this project aims at understanding and predicting extreme events of energy dissipation in atmospheric turbulence, by:
(1) advanced measurements of extreme events in a simplified model experiment,
(2) the comparison with numerical simulations,
(3) the development of ML data-driven models from experimental data.
This project aims at understanding and predicting extreme events of energy dissipation from experiments in Rayleigh-Bénard convection (RBC), a turbulent fluid layer that is uniformly heated from below and cooled from above. RBC is considered as a paradigm for atmospheric turbulence. Measurements of extreme events of energy dissipation are performed using Particle Image Velocimetry (PIV).
Machine Learning (ML) methods are an efficient and unbiased way to process a large amount of data effectively. ML has been rarely applied to the detection of extreme events in fluid turbulence. In this project, by using Recurrent Neural Networks (a particular king of ML method), a large amount of data from experiments and simulations is processed to construct a model. The predicted extreme events are large wind fluctuations at small scales, wind gusts, which may have important consequences for wind turbines operating conditions.
In summary, this project aims at understanding and predicting extreme events of energy dissipation in atmospheric turbulence, by:
(1) advanced measurements of extreme events in a simplified model experiment,
(2) the comparison with numerical simulations,
(3) the development of ML data-driven models from experimental data.
Three-dimensional (3D) instantaneous velocity fields were measured in a RBC cell filled with (compressed) air in flows from weakly turbulent to fully turbulent. The experimental technique used to acquire the data is stereoscopic PIV. The RBC cell set-up was inserted in the SCALEX facility of TU Ilmenau (Germany), a pressure vessel with several optical accesses that can be pressurized up to 10 bars. To study turbulent flows, the whole experimental facility, including cameras and objective lenses, was pressurized up to 4.5 bars. The main goal of the first part of the project was to reproduce Direct Numerical Simulations (DNS) data of the same flow [1]. Both the experimental data and the DNS ones are used to study extreme events of vorticity. The experimental results were able to reproduce the statistics of DNS data of the same flow well and allowed the study of extreme events of vorticity. From Probability Density Functions (PDFs) statistics both from PIV experiments and from DNS data we observed a transition to intermittency in the turbulent regime. These results were presented at the International Symposium of PIV in August 2021 [3].
In the second part of the project, the experimental investigation was connected with a comprehensive statistical analysis of long-term time series of vorticity and individual velocity derivatives. A statistical convergence for derivative moments up to an order of six is demonstrated. Our results are found to agree well with existing high-resolution direct numerical simulation data in the same range of parameters, including the extreme vorticity events that appear in the far exponential tails of the corresponding probability density functions [2]. The experimental data were used to train a reservoir computing model, one implementation of a RNN, to reproduce highly intermittent experimental time series of the vorticity and thus reconstruct extreme out-of-plane vorticity events. After training the model with high-resolution PIV data, the machine learning model is run with sparsely seeded, continually available, and unseen measurement data in the reconstruction phase. The dependence of the reconstruction quality on the sparsity of the partial observations is also documented [2]. Our latter result paves the way to machine-learning-assisted experimental analyses of small-scale turbulence for which time series of missing velocity derivatives can be provided by generative algorithms. These results were presented at three conferences: (1) the Annual meeting of the American Physical society, Division of Fluid Dynamics, on November 2021, (2) the annual meeting of the priority program SPP 1881 funded by the German Research Foundation on June 2022, and (3) the EuroMech colloquium 619 on July 2022. They are published in [1-3].
[1] Valentina Valori and Jörg Schumacher, Europhysics Letters, 2021. DOI: 10.1209/0295-5075/134/34004
[2] Valentina Valori, Robert Kräuter and Jörg Schumacher, Physical Review Research, 2022, DOI: https://doi.org/10.1103/PhysRevResearch.4.023180(se abrirá en una nueva ventana)
[3] Valori Valentina, Alexander Thieme, Christian Cierpka, Joerg Schumacher, 14th International Symposium on Particle Image Velocimetry, ISPIV 2021
In the second part of the project, the experimental investigation was connected with a comprehensive statistical analysis of long-term time series of vorticity and individual velocity derivatives. A statistical convergence for derivative moments up to an order of six is demonstrated. Our results are found to agree well with existing high-resolution direct numerical simulation data in the same range of parameters, including the extreme vorticity events that appear in the far exponential tails of the corresponding probability density functions [2]. The experimental data were used to train a reservoir computing model, one implementation of a RNN, to reproduce highly intermittent experimental time series of the vorticity and thus reconstruct extreme out-of-plane vorticity events. After training the model with high-resolution PIV data, the machine learning model is run with sparsely seeded, continually available, and unseen measurement data in the reconstruction phase. The dependence of the reconstruction quality on the sparsity of the partial observations is also documented [2]. Our latter result paves the way to machine-learning-assisted experimental analyses of small-scale turbulence for which time series of missing velocity derivatives can be provided by generative algorithms. These results were presented at three conferences: (1) the Annual meeting of the American Physical society, Division of Fluid Dynamics, on November 2021, (2) the annual meeting of the priority program SPP 1881 funded by the German Research Foundation on June 2022, and (3) the EuroMech colloquium 619 on July 2022. They are published in [1-3].
[1] Valentina Valori and Jörg Schumacher, Europhysics Letters, 2021. DOI: 10.1209/0295-5075/134/34004
[2] Valentina Valori, Robert Kräuter and Jörg Schumacher, Physical Review Research, 2022, DOI: https://doi.org/10.1103/PhysRevResearch.4.023180(se abrirá en una nueva ventana)
[3] Valori Valentina, Alexander Thieme, Christian Cierpka, Joerg Schumacher, 14th International Symposium on Particle Image Velocimetry, ISPIV 2021
From the experimental campaign it was experimentally observed for the first time a transition from Gaussian to non-Gaussian velocity derivative statistics in the bulk of a convection flow. This result was know only from DNS data before. Therefore, the acquired PIV measurements experimentally confirmed the DNS data.
A reservoir computing model, a particular kind of RNN was developed trained on the long time series experiments. This model was able to reproduce the highly intermittent experimental time series of the vorticity, and to reconstruct extreme events of out-of-plane vorticity.
In summary the three scientific objectives of the project were successfully met. Indeed: Advanced measurements of the three dimensional velocity field were performed to study extreme events in thermal convection. These measurements confirmed existing DNS data. The measurements were used to train a newly developed ML algorithm that is able to reconstruct extreme events of vorticity.
A reservoir computing model, a particular kind of RNN was developed trained on the long time series experiments. This model was able to reproduce the highly intermittent experimental time series of the vorticity, and to reconstruct extreme events of out-of-plane vorticity.
In summary the three scientific objectives of the project were successfully met. Indeed: Advanced measurements of the three dimensional velocity field were performed to study extreme events in thermal convection. These measurements confirmed existing DNS data. The measurements were used to train a newly developed ML algorithm that is able to reconstruct extreme events of vorticity.