Periodic Reporting for period 2 - HIFI-TURB (HIGH-FIDELITY LES/DNS DATA FOR INNOVATIVE TURBULENCE MODELS)
Okres sprawozdawczy: 2021-01-01 do 2022-12-31
The most significant challenge in all areas of applied fluid dynamics is posed by a lack of understanding and thus poor prediction of turbulence dependent features. Improving the capabilities of turbulence models when predicting these complex or separated fluid flows offers the potential of reducing energy consumption of aircraft, cars, and ships, with consequent reduction in emissions and noise.
The HiFi-TURB project sets out a highly ambitious and innovative program of work designed to address influential deficiencies in advanced statistical models of turbulence. Current industrial practice relies greatly on turbulence modelling implemented within the Reynolds-Averaged Navier-Stokes (RANS) framework, wherein turbulence is described by models represented by ensemble-averaged properties.
Time-resolved Large Eddy Simulations (LES) and Direct Numerical Simulation (DNS) yield superior realism in representing turbulence, but they are not directly applicable in the most of the industrial design work due to the high computational cost and they generate massive data sets that require elaborate and time-consuming statistical analysis.
The vision – the paradigm – of the HiFi-TURB project is to generate and exploit LES- and DNS- data for a carefully selected set of flow configurations that contain, collectively, most features of interest of complex 3D flows and separated regions. This is done for the purpose of improving substantially advanced RANS models for industrial use. At the core of the project, various Artificial Intelligence (AI) and Machine Learning (ML) techniques are applied to process the large amounts of data generated and to gain new insights into the physics of such flows. It is part of the rationale of the project that ML driven modelling approaches are guided by world leading experts in turbulence modelling, namely F. Menter, P. Spalart, M. Leschziner, S. Wallin, S. Jakirlic, W. Rodi, M.V. Salvetti.
The generated database and model results are disseminated towards the wide research community, via the ERCOFTAC Knowledge Base Wiki.
HiFi-TURB can claim to have been a trail-blazing learning exercise and significant technical progress is achieved in all work packages. The project was characterized by a high level of interaction between the partners involved in the various work packages. The validity of the HiFi-TURB paradigm has been successfully demonstrated, by deriving interpretable turbulence models from high-fidelity statistical data and by implementing them successfully into several simulation codes.
The HiFi-TURB project sets out a highly ambitious and innovative program of work designed to address influential deficiencies in advanced statistical models of turbulence. Current industrial practice relies greatly on turbulence modelling implemented within the Reynolds-Averaged Navier-Stokes (RANS) framework, wherein turbulence is described by models represented by ensemble-averaged properties.
Time-resolved Large Eddy Simulations (LES) and Direct Numerical Simulation (DNS) yield superior realism in representing turbulence, but they are not directly applicable in the most of the industrial design work due to the high computational cost and they generate massive data sets that require elaborate and time-consuming statistical analysis.
The vision – the paradigm – of the HiFi-TURB project is to generate and exploit LES- and DNS- data for a carefully selected set of flow configurations that contain, collectively, most features of interest of complex 3D flows and separated regions. This is done for the purpose of improving substantially advanced RANS models for industrial use. At the core of the project, various Artificial Intelligence (AI) and Machine Learning (ML) techniques are applied to process the large amounts of data generated and to gain new insights into the physics of such flows. It is part of the rationale of the project that ML driven modelling approaches are guided by world leading experts in turbulence modelling, namely F. Menter, P. Spalart, M. Leschziner, S. Wallin, S. Jakirlic, W. Rodi, M.V. Salvetti.
The generated database and model results are disseminated towards the wide research community, via the ERCOFTAC Knowledge Base Wiki.
HiFi-TURB can claim to have been a trail-blazing learning exercise and significant technical progress is achieved in all work packages. The project was characterized by a high level of interaction between the partners involved in the various work packages. The validity of the HiFi-TURB paradigm has been successfully demonstrated, by deriving interpretable turbulence models from high-fidelity statistical data and by implementing them successfully into several simulation codes.
In WP 2 robust and scalable adaptive methods are proposed to guarantee a proper spatial and temporal resolution while reducing the cost of simulations. The work focused on exploiting the most recent computer hardware based on GPUs. New algorithms and functionalities for curved mesh generation are included in production tools and tested on industrial configurations.
In WP 3 large computational resources granted from PRACE and EuroHPC calls were used by High-Order solvers to generate detailed high-fidelity reference. Two datasets have already been made publicly available in the ERCOFTAC Knowledge Base Wiki, a third one is in preparation for publication while other databases are still being generated beyond the end of the project.
In WP 4 machine learning methodologies were developed with the goal to exploit high-fidelity datasets for the purpose of turbulence model improvements. Several machine learning techniques are developed and successfully applied based on a novel framework for turbulence modelling proposed in WP 5. It is demonstrated that the ML approaches can lead to improved turbulence models by learning from suitable high-fidelity data as defined within the HiFi-TURB project.
In WP 5 thanks to substantial efforts by the project partners and a tight collaboration between WP 5 and WP 4, new model EARSM and DRSM forms have emerged. Data-driven approaches have also been explored in scale-resolving LES and hybrid RANS-LES methods – specifically, to optimize the manner in which the LES solution and the wall is bridged through coupling the outer solution to the near-wall RANS model. These efforts have identified some promising directions along which to continue pursuing ML-based modelling. However, it must be acknowledged that well-posed general models, demonstrably applicable to a wide range of conditions, have yet to emerge.
In WP 6 simulations with ML-based DRSM and EARSM models developed during the HiFi-TURB project were applied to seven test cases of increasing complexity. Due to the interpretability of the produced models, different groups were able to implement and test them. In some cases, the new models were found to improve the prediction of key quantities, in other cases the results produced were close to existing baseline models.
In WP 7 the existing ERCOFTAC Knowledge Base Wiki has been extended for storing the DNS data obtained in WP 3 with accompanying documentation. Lists of desirable statistical quantities to be stored, a template for entering the data and documentation, and guidelines for data storage were developed. Data and documentation were implemented in the Wiki for four test cases produced during the project. The information is openly accessible and hence of use to the entire fluid mechanics community and should have an impact on the further improvement of prediction methods.
In WP 3 large computational resources granted from PRACE and EuroHPC calls were used by High-Order solvers to generate detailed high-fidelity reference. Two datasets have already been made publicly available in the ERCOFTAC Knowledge Base Wiki, a third one is in preparation for publication while other databases are still being generated beyond the end of the project.
In WP 4 machine learning methodologies were developed with the goal to exploit high-fidelity datasets for the purpose of turbulence model improvements. Several machine learning techniques are developed and successfully applied based on a novel framework for turbulence modelling proposed in WP 5. It is demonstrated that the ML approaches can lead to improved turbulence models by learning from suitable high-fidelity data as defined within the HiFi-TURB project.
In WP 5 thanks to substantial efforts by the project partners and a tight collaboration between WP 5 and WP 4, new model EARSM and DRSM forms have emerged. Data-driven approaches have also been explored in scale-resolving LES and hybrid RANS-LES methods – specifically, to optimize the manner in which the LES solution and the wall is bridged through coupling the outer solution to the near-wall RANS model. These efforts have identified some promising directions along which to continue pursuing ML-based modelling. However, it must be acknowledged that well-posed general models, demonstrably applicable to a wide range of conditions, have yet to emerge.
In WP 6 simulations with ML-based DRSM and EARSM models developed during the HiFi-TURB project were applied to seven test cases of increasing complexity. Due to the interpretability of the produced models, different groups were able to implement and test them. In some cases, the new models were found to improve the prediction of key quantities, in other cases the results produced were close to existing baseline models.
In WP 7 the existing ERCOFTAC Knowledge Base Wiki has been extended for storing the DNS data obtained in WP 3 with accompanying documentation. Lists of desirable statistical quantities to be stored, a template for entering the data and documentation, and guidelines for data storage were developed. Data and documentation were implemented in the Wiki for four test cases produced during the project. The information is openly accessible and hence of use to the entire fluid mechanics community and should have an impact on the further improvement of prediction methods.
Progress beyond the state-of-the art has been achieved across all work packages of the HiFi-TURB project. This includes:
• Significant simulation code improvements that saved millions of CPU core hours during the generation of reference data within the project
• The definition of a list of desirable and minimal statistical and time series quantities which should be collected during high-fidelity flow simulations to investigate new turbulence models through machine learning. This list is made available via the ERCOFTAC Wiki and will serve the turbulence modelling efforts far beyond the HiFi-TURB project.
• A major novel framework for turbulence modelling in view of the specific challenges of a ML based approach for EARSM. It preserves realizability, improves physical interpretation and is well-posed for the ML task.
• Successful demonstration of the applicability of several ML approaches to the task of turbulence model improvement, leading to several new turbulence model variations learnt from high-fidelity data.
• Interpretability of the ML produced models which allows for a wide dissemination, as proven by the implementation and testing of those models by different partners in their respective simulation codes
• Test case data and documentation have been implemented in the ERCOFTAC Wiki for four test cases produced during the HiFi-TURB project respecting FAIR data conditions. The data stored provide rich information beyond the state of the art on a set of fairly complex flows with relevance in engineering applications
The results of the HiFi-TURB project will be the building blocks for continued efforts in turbulence modelling in the EU and beyond and the best practices, standards and data generated within the project will serve the research community. The project will as such have a major impact on environmental factors helping to leverage the potential of reducing energy consumption of aircraft, cars, and ships, with consequent reduction in emissions and noise.
• Significant simulation code improvements that saved millions of CPU core hours during the generation of reference data within the project
• The definition of a list of desirable and minimal statistical and time series quantities which should be collected during high-fidelity flow simulations to investigate new turbulence models through machine learning. This list is made available via the ERCOFTAC Wiki and will serve the turbulence modelling efforts far beyond the HiFi-TURB project.
• A major novel framework for turbulence modelling in view of the specific challenges of a ML based approach for EARSM. It preserves realizability, improves physical interpretation and is well-posed for the ML task.
• Successful demonstration of the applicability of several ML approaches to the task of turbulence model improvement, leading to several new turbulence model variations learnt from high-fidelity data.
• Interpretability of the ML produced models which allows for a wide dissemination, as proven by the implementation and testing of those models by different partners in their respective simulation codes
• Test case data and documentation have been implemented in the ERCOFTAC Wiki for four test cases produced during the HiFi-TURB project respecting FAIR data conditions. The data stored provide rich information beyond the state of the art on a set of fairly complex flows with relevance in engineering applications
The results of the HiFi-TURB project will be the building blocks for continued efforts in turbulence modelling in the EU and beyond and the best practices, standards and data generated within the project will serve the research community. The project will as such have a major impact on environmental factors helping to leverage the potential of reducing energy consumption of aircraft, cars, and ships, with consequent reduction in emissions and noise.