One of the main concerns of the current scientific community is to be able to understand the underlying physics involved in complex aerodynamic flow configurations. This is a challenge since current constraints in hydrogen propulsion or less emissions of pollutants is driving aviation to new schemes where industry still does not feel comfortable. In order to understand those configurations, numerical simulations or experiments play a key role in predicting flow behaviour. However, those studies generate a huge amount of data that makes it difficult to manipulate and extract useful and valuable information for industry. This is the role of big data analysis.

Big data analysis is not only limited to visualisation. In the NextSim project, huge amounts of data will be generated through several very challenging industrial test cases. Most of them result in unsteadiness, where the flow evolves in time and space in a very complex manner. Extracting information from this huge amount of data is becoming difficult but also critical to accurately understand the relevant physical features of the problem. The NextSim project is investigating and developing several feature detection techniques, such as (Spectral) Proper Orthogonal or (High Order) Dynamic mode decomposition, in a way that will allow them to become regular tools in the analysis of fluid flows. NextSim is working to mature those tools to be able to provide information in huge and non-uniform databases, or when the number of scales involved is very large, such as turbulent flows.

In more detail, Work Package (WP) 3 of NextSim: Algorithms for Data Management, Visualization and Modelling, proposes to facilitate the post-process and analysis of the large amount of data generated during the simulations.

This WP, led by UPM with the support of BSC, CIMNE, ONERA and CERFACS, aims to couple CODA outputs with feature (stability) analysis tool, detection techniques (e.g. POD/DMD) and machine learning methods to detect the most flow features. Moreover the tasks in this WP include:

  1. Generating reduced order models (ROM) than can mimic the complex flow and be used interpolate and extrapolate data for a variety of flow conditions.
  2. Developing visualization strategies for HO methods working in large distributed memory clusters for performance.
  3. Developing optimisation loops to accelerate the finding of optimised aircraft geometries

Reported activities during this first year are related to the development of a scalable library which includes several Low Order Methods and feature detection algorithms for data analysis, in particular:

  • Dynamic Mode Decomposition (DMD) – computes a set of modes each of which is associated with a fixed oscillation frequency and decay/growth rate. These modes are approximations of the modes and eigenvalues of the Koopman operator.
  • High Order DMD (HODMD) – A more sophisticated DMD method. It was developed to work with complex flows, i.e., transitions to turbulence, feature detection in noisy data.
  • Resolvent analysis (RA) identifies the most responsive forcing and most receptive states of a dynamical system, in an input–output sense, based on its governing equations.
  • Surrogate models based on Singular Value Decomposition which includes randomize SVD or HOSVD for efficient interpolation and surrogate modelling in several databases.

Preliminary results have been obtained in the High-Lift CRM configuration free-air, with Angle of Attach of 19 degrees and Reynolds number of 5.5Million (https://hiliftpw.larc.nasa.gov/Workshop4/WorkshopPresentations/07_GMGW3_HLPW4_WMLES-LB.pdf).  The data has been obtained from a numerical simulation obtained with a wmLES, FE low dissipation (order 2), LES closure SGS Vreman. The mesh is composed of 273M elements, 112 snapshots with about 300Gb of data to process.

As a reference, the time to carry out the POD analysis 5 min with 2000 cpu’s. For illustration, a snapshot of the computational solutions and a section of the first POD mode are shown below.

A section of the first POD mode