Periodic Reporting for period 1 - HIFI-TURB (HIGH-FIDELITY LES/DNS DATA FOR INNOVATIVE TURBULENCE MODELS)
Période du rapport: 2019-07-01 au 2020-12-31
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 and in particular separated fluid flows offers the potential of reducing energy consumption of aircraft, cars, and ships, with consequent reduction in emissions and noise. And has as such a major impact on environmental factors as well as on economy and industrial leadership in the highly competitive global markets.
Against this background, the HiFi-TURB project sets out a highly ambitious and innovative programme 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 exploit LES- and DNS-generated data for a carefully selected set of flow configurations that contain, collectively, most features of interest of complex 3D flows and separated regions, for the purpose of improving substantially advanced RANS models for industrial use. An ambitious quantitative objective set by HiFi-TURB is the aim of an error margin below 10% in operationally influential global properties in separation-dominated flow properties - for example, the maximum lift prediction.
The methodology to be followed combines the following elements:
• The development of new modelling approaches focused exclusively on anisotropy resolving differential and algebraic Reynolds-stress models (rather than linear eddy-viscosity models)
• The generation of high-fidelity DNS and WRLES data representative of different separation inducing flow configurations, which can be achieved with the present Petascale HPC capabilities, with high order methods.
• The exploitation of various Artificial Intelligence and Machine Learning techniques, to process the large amounts of data generated and to gain potentially new insights into the physics of such flows. This essential approach is highly innovative and will require prospecting different roads and AI techniques, with objectives being guided and monitored by the turbulence expert group
• The presence in the consortium of the worldwide top level experts in turbulence modelling, namely F. Menter, P. Spalart, L. Leschziner, S. Wallin, S. Jakirlic, W. Rodi guiding the investigations towards new modelling ideas
• The database dissemination towards the wide research community, via the ERCOFTAC Knowledge Base Wiki
Against this background, the HiFi-TURB project sets out a highly ambitious and innovative programme 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 exploit LES- and DNS-generated data for a carefully selected set of flow configurations that contain, collectively, most features of interest of complex 3D flows and separated regions, for the purpose of improving substantially advanced RANS models for industrial use. An ambitious quantitative objective set by HiFi-TURB is the aim of an error margin below 10% in operationally influential global properties in separation-dominated flow properties - for example, the maximum lift prediction.
The methodology to be followed combines the following elements:
• The development of new modelling approaches focused exclusively on anisotropy resolving differential and algebraic Reynolds-stress models (rather than linear eddy-viscosity models)
• The generation of high-fidelity DNS and WRLES data representative of different separation inducing flow configurations, which can be achieved with the present Petascale HPC capabilities, with high order methods.
• The exploitation of various Artificial Intelligence and Machine Learning techniques, to process the large amounts of data generated and to gain potentially new insights into the physics of such flows. This essential approach is highly innovative and will require prospecting different roads and AI techniques, with objectives being guided and monitored by the turbulence expert group
• The presence in the consortium of the worldwide top level experts in turbulence modelling, namely F. Menter, P. Spalart, L. Leschziner, S. Wallin, S. Jakirlic, W. Rodi guiding the investigations towards new modelling ideas
• The database dissemination towards the wide research community, via the ERCOFTAC Knowledge Base Wiki
The first activities concentrated on the development of High-Order Methods (HOM) and curved mesh generation to reduce the computational cost of scale-resolving simulations. Lines of research include adaptive methods for mesh and local polynomial degree adaptation, implicit methods, smart initialization procedures and GPU-acceleration of HO solvers. Numerical tests on first flow problems have shown significant savings in CPU time.
To align with the requirements of HOM, partners also pursued research on the generation of curved grid generation focusing mainly on the boundary layer spacing and curving.
These HO solvers will be used to establish detailed high-fidelity reference databases that can be used for the turbulence modelling development. These databases will be made publicly available on the ERCOFTAC Knowledge Base Wiki. With regard to the database, first, the test cases were carefully defined with fully known and reproducible conditions. A procedure to inject a turbulent boundary layer with the desired characteristics inspired from the available literature has been proposed. This is to ensure that the boundary layer has reached a canonical state at a given reference location upstream the region of interest with well established characteristics.
Second, a list of desirable and minimum statistical and time series quantities which need to be collected during the computations has been determined based on discussions with experts in turbulence modeling. And third, the infrastructure to store the statistical data has been put into place, ready to receive the data that will be prepared during the second half of the project.
The HiFi-TURB project includes a turbulence modelling expert group whose role is to oversee all fundamental turbulence-modelling efforts in the project. The objectives are (i) to formulate and/or define baseline models that form the starting point of machine-learning improvements in WP4; (ii) to guide the evolution of improved models through continuous interaction with WP4, (iii) to evaluate proposals arising from WP4, and to oversee the verification and validation of emerging models by reference to test cases.
Guided by the turbulence modelling expert group, various ML approaches have been applied to the challenge of gaining new insights into the physics of turbulence, and to further improve and to develop new and physically sound Reynolds stress turbulence models. Various approaches such as Variational Auto-Encoders, Gene Expression Programming or Multiple Expression programming are pursued, showing some first success.
The consortium has prepared solutions for defined test cases with currently existing turbulence models to prepare a benchmark for the new models under development within the project.
To align with the requirements of HOM, partners also pursued research on the generation of curved grid generation focusing mainly on the boundary layer spacing and curving.
These HO solvers will be used to establish detailed high-fidelity reference databases that can be used for the turbulence modelling development. These databases will be made publicly available on the ERCOFTAC Knowledge Base Wiki. With regard to the database, first, the test cases were carefully defined with fully known and reproducible conditions. A procedure to inject a turbulent boundary layer with the desired characteristics inspired from the available literature has been proposed. This is to ensure that the boundary layer has reached a canonical state at a given reference location upstream the region of interest with well established characteristics.
Second, a list of desirable and minimum statistical and time series quantities which need to be collected during the computations has been determined based on discussions with experts in turbulence modeling. And third, the infrastructure to store the statistical data has been put into place, ready to receive the data that will be prepared during the second half of the project.
The HiFi-TURB project includes a turbulence modelling expert group whose role is to oversee all fundamental turbulence-modelling efforts in the project. The objectives are (i) to formulate and/or define baseline models that form the starting point of machine-learning improvements in WP4; (ii) to guide the evolution of improved models through continuous interaction with WP4, (iii) to evaluate proposals arising from WP4, and to oversee the verification and validation of emerging models by reference to test cases.
Guided by the turbulence modelling expert group, various ML approaches have been applied to the challenge of gaining new insights into the physics of turbulence, and to further improve and to develop new and physically sound Reynolds stress turbulence models. Various approaches such as Variational Auto-Encoders, Gene Expression Programming or Multiple Expression programming are pursued, showing some first success.
The consortium has prepared solutions for defined test cases with currently existing turbulence models to prepare a benchmark for the new models under development within the project.
An overall improvement of the computational workflow for High-Order Methods (HOM) based CFD solvers, including the curved mesh generation phase has been reached. Significant progress has been reached by introduction of the above mentioned elements.
Mesh generation is often considered as a bottleneck in the CFD workflow, especially when dealing with HO solvers that need curved grids. Automatic high-order meshing capabilities that meet the accuracy constraints of CFD simulations will significantly alleviated the costs of the grid generation phase.
Progress beyond state of the art accomplished so far includes the definition of a list of desirable and minimal statistical and time series quantities which should be collected during high-fidelity flow simulations in order to investigate new turbulence models through machine learning. Along this list, the best practices to ensure the quality of the accumulated statistics have also been investigated and reported.
A major novel approach is proposed in the context of EARSM models by Stefan Wallin (Ercoftac) in view of the specific challenges of the ML based approach to turbulence modelling. The introduction of a novel scaling method makes tensor terms and their invariants independent from the dissipation rate and realizability can be more easily preserved. This represents a significant step beyond the state-of-the-art.
In general, the interactions between the turbulence-modelling and machine-learning activities go well beyond the present state of the art and are effective in laying the foundation for more substantive efforts to be pursued over the second half of the project.
Mesh generation is often considered as a bottleneck in the CFD workflow, especially when dealing with HO solvers that need curved grids. Automatic high-order meshing capabilities that meet the accuracy constraints of CFD simulations will significantly alleviated the costs of the grid generation phase.
Progress beyond state of the art accomplished so far includes the definition of a list of desirable and minimal statistical and time series quantities which should be collected during high-fidelity flow simulations in order to investigate new turbulence models through machine learning. Along this list, the best practices to ensure the quality of the accumulated statistics have also been investigated and reported.
A major novel approach is proposed in the context of EARSM models by Stefan Wallin (Ercoftac) in view of the specific challenges of the ML based approach to turbulence modelling. The introduction of a novel scaling method makes tensor terms and their invariants independent from the dissipation rate and realizability can be more easily preserved. This represents a significant step beyond the state-of-the-art.
In general, the interactions between the turbulence-modelling and machine-learning activities go well beyond the present state of the art and are effective in laying the foundation for more substantive efforts to be pursued over the second half of the project.