Periodic Reporting for period 2 - NEXTFLOW (Next-generation flow diagnostics for control)
Reporting period: 2022-07-01 to 2023-12-31
The vast majority of flows in nature and man-made applications is turbulent, i.e. with chaotic and apparently irregular behaviour. The impact of turbulent flows in our life is overwhelming. For instance, the drag of all transport vehicles, and consequently fuel consumption and emissions, depend strongly on the behaviour of turbulent flows. Understanding and learning how to control them has a potentially tremendous impact on the sustainability of human development. Measuring and controlling turbulent flows represents, however, a formidable challenge. Their chaotic three-dimensional multi-scale character requires measurement techniques capable of extracting complete high-fidelity detailed views of their behaviour. In most cases, however, current measurement techniques are able to provide only partial views, for instance only of velocity or thermodynamic quantities, and often only in point or planar domains, or with loss in resolution for volumetric measurements.
The project NEXTFLOW aims at developing the next-generation flow-measurement techniques for turbulent flows by augmenting existing methods with data-driven tools. The central hypothesis is that combining multiple measurement techniques, each one providing a different incomplete view, and physical principles, makes it possible to disclose a complete description of the fluid flow. We build upon three-dimensional particle image velocimetry and extend its capabilities beyond its limits in temporal and spatial resolution by blending in the information of fast probes and including constraints based on flow physics. The final objective is to exploit the complete flow description to distil compact models, which can be used for flow control applications.
The project NEXTFLOW aims at developing the next-generation flow-measurement techniques for turbulent flows by augmenting existing methods with data-driven tools. The central hypothesis is that combining multiple measurement techniques, each one providing a different incomplete view, and physical principles, makes it possible to disclose a complete description of the fluid flow. We build upon three-dimensional particle image velocimetry and extend its capabilities beyond its limits in temporal and spatial resolution by blending in the information of fast probes and including constraints based on flow physics. The final objective is to exploit the complete flow description to distil compact models, which can be used for flow control applications.
In the first part of the project, the NEXTFLOW team achieved the following results:
- A data-driven method to obtain time-resolved pressure fields from non-time-resolved velocity fields and high-repetition-rate probe measurements was developed. The method builds upon the Extended Proper Orthogonal Decomposition (EPOD) to obtain time-resolved velocity fields and enforces the Navier-Stokes equation to compute the pressure gradient and, from integration, the pressure fields. Methods to increase the accuracy by using neural networks and introducing physical constraints have also been explored. This is a key enabler for complete flow measurements in cases where models cannot be included, either because the data are largely incomplete (for instance planar data) or because the assumptions underlying the models are not valid.
- We developed a 3D generative adversarial network (GAN) to estimate 3D velocity fields using wall-shear and pressure data.
- We proposed a novel end-to-end tool for spatial resolution enhancement and uncertainty quantification based on dimensionality reduction on a local basis, and K-nearest neighbours (KNN) search to merge information from topologically-similar (although statistically-independent) snapshots. The method achieves a substantial improvement in accuracy and spatial resolution with respect to state-of-art methods, and it provides the benefit of uncertainty quantification directly embedded in the process. The estimation of the uncertainty is of special relevance for validation benchmarks and data assimilation.
- A concept blending our KNN method with Radial Basis Function for the extension to 3D was explored. The method allows the introduction of physical constraints and the analytical derivation of pressure fields and low-order models.
- An algorithm for super-resolution based on GANs using solely sparse data was demonstrated. While standard methods require pre-trained networks or high-resolution dictionaries for training, we delivered a new approach that relies solely on the sparse data for training and does not require any labelled dataset within the process. The technique is very robust to noise and demonstrated unprecedented levels of accuracy.
- Active flow control experiments optimized with genetic algorithms were successfully carried out.
- A jet facility to perform active flow control experiments in a controlled environment and under a range of different operating conditions was designed.
- An intense dissemination activity targeted to maximize impact on both the scientific community and the general public was carried out. In addition to standard venues (publications in well-renowned journals, presentations at international conferences, seminars, workshops), we published open-access databases and codes whenever applicable to maximize impact and allow straightforward extension of our work. Furthermore, we proposed a large number of activities for people of all ages in events such as the Fair of Science, the Week of Science, and the European Researchers’ Night.
- A data-driven method to obtain time-resolved pressure fields from non-time-resolved velocity fields and high-repetition-rate probe measurements was developed. The method builds upon the Extended Proper Orthogonal Decomposition (EPOD) to obtain time-resolved velocity fields and enforces the Navier-Stokes equation to compute the pressure gradient and, from integration, the pressure fields. Methods to increase the accuracy by using neural networks and introducing physical constraints have also been explored. This is a key enabler for complete flow measurements in cases where models cannot be included, either because the data are largely incomplete (for instance planar data) or because the assumptions underlying the models are not valid.
- We developed a 3D generative adversarial network (GAN) to estimate 3D velocity fields using wall-shear and pressure data.
- We proposed a novel end-to-end tool for spatial resolution enhancement and uncertainty quantification based on dimensionality reduction on a local basis, and K-nearest neighbours (KNN) search to merge information from topologically-similar (although statistically-independent) snapshots. The method achieves a substantial improvement in accuracy and spatial resolution with respect to state-of-art methods, and it provides the benefit of uncertainty quantification directly embedded in the process. The estimation of the uncertainty is of special relevance for validation benchmarks and data assimilation.
- A concept blending our KNN method with Radial Basis Function for the extension to 3D was explored. The method allows the introduction of physical constraints and the analytical derivation of pressure fields and low-order models.
- An algorithm for super-resolution based on GANs using solely sparse data was demonstrated. While standard methods require pre-trained networks or high-resolution dictionaries for training, we delivered a new approach that relies solely on the sparse data for training and does not require any labelled dataset within the process. The technique is very robust to noise and demonstrated unprecedented levels of accuracy.
- Active flow control experiments optimized with genetic algorithms were successfully carried out.
- A jet facility to perform active flow control experiments in a controlled environment and under a range of different operating conditions was designed.
- An intense dissemination activity targeted to maximize impact on both the scientific community and the general public was carried out. In addition to standard venues (publications in well-renowned journals, presentations at international conferences, seminars, workshops), we published open-access databases and codes whenever applicable to maximize impact and allow straightforward extension of our work. Furthermore, we proposed a large number of activities for people of all ages in events such as the Fair of Science, the Week of Science, and the European Researchers’ Night.
NEXTFLOW opened the way to time-resolved pressure measurements based on velocity fields estimated using the simultaneous acquisition of non-time-resolved velocity fields and fast probes. Unlike existing methods, it does not require modeling assumptions, nor necessarily 3D data to build upon, thus advancing the current state of the art. Novel advances leveraging physical constraints to regularize the data and extend the capabilities of the method are envisaged.
Furthermore, the project has delivered techniques to increase the spatial resolution of non-time-resolved Particle Image Velocimetry (PIV) to an unprecedented level. Considering the cost and technical limitations of high-repetition-rate equipment, both advances in time and space resolution are building the ground for the use of the consolidated (and widely available in laboratories) low-repetition-rate PIV to perform accurate time-resolved velocity and pressure measurements.
The availability of time-resolved data using cost-effective equipment paves the way to more accessible measurements of the flow dynamics. This is expected to have an impact on flow control applications, in which understanding the dynamics of fluid mechanical devices with and without control is essential to optimize their performances.
Furthermore, the project has delivered techniques to increase the spatial resolution of non-time-resolved Particle Image Velocimetry (PIV) to an unprecedented level. Considering the cost and technical limitations of high-repetition-rate equipment, both advances in time and space resolution are building the ground for the use of the consolidated (and widely available in laboratories) low-repetition-rate PIV to perform accurate time-resolved velocity and pressure measurements.
The availability of time-resolved data using cost-effective equipment paves the way to more accessible measurements of the flow dynamics. This is expected to have an impact on flow control applications, in which understanding the dynamics of fluid mechanical devices with and without control is essential to optimize their performances.