Self-guided Machine Learning Algorithms for Real-Time Assimilation, Interpolation and Rendering of Flow Data
The proposed project addresses several major challenges encountered in the assimilation of measurement data in aerodynamic testing.
Optimization plays a central role in the design of current and future transportation systems such as trains, airplanes or automobiles. The objective is to develop designs with reduced energy consumption, smaller environmental footprint and increased customer comfort. For this, numerical simulations and experimental tests are being performed, both with their own set of constraints such as turnover times and cost. The Institute of Fluid Dynamics (IFD) operates a wind tunnel for such testing, and the research is focused on the development of novel, efficient measurement techniques to enhance the science data return from such cost-intensive facilities.
The proposed project will focus on two central problems limiting the productivity of experimental aerodynamic test campaigns. Strategies using machine learning will be developed to dynamically assimilate acquired data into a global description of the flow field being measured. Predictive analysis of the data will be employed to direct the measurements process towards regions of significant information as the global knowledge evolves. The measurement time will be reduced relying on adaptive guidance based on real-time data interpretation. The software design will explicitly include the option of a human operator in the loop.
The collaboration between SDSC and IFD offers an attractive way to merge complementary competencies. The tasks of aerodynamic flow field reconstruction, uncertainty quantification and probe guidance will be broken down into distinct activities / work packages such as (1) development of learning algorithms for sparse flow data assimilation using physics-based constraint models, (2) real-time implementation of the software in a suitable computing infrastructure and (3) testing and evaluation of the complete hardware/software system in situ in the wind tunnel facility at IFD.
Measuring the flow field surrounding an object using the probe system gives a continuous stream of data. Based on this data, the goal of this interactive procedure is the volumetric reconstruction of the
mean properties (flow direction, magnitude) and derived quantities (vorticity, stream-lines, etc.) of the flow field with the highest fidelity in minimal time. To do so, we aim at developing a machine learning method that can reconstruct the flow field based on the incoming data. In addition, we aim at reducing the experiment time using a machine learning tool identifying significant regions of interest where further acquisition will help to reduce the uncertainties and improve the information about the flow field.
the real-time capability of the algorithm as well as its adaptive aspect will give an optimized probe trajectory. It will provide improvements on the real-time signal reconstruction process.
The SDSC will lead the WPs related with the development of the machine learning tools. First, the SDSC will help to improve the current flow field reconstruction algorithm. Second, the SDSC will develop a generic real-time algorithm that targets the location where sampling might reduce the uncertainty about the flow field. Third, it will contribute to the implementation of the algorithm in the guided probe software.
Probe trajectory around the object. The red arrows correspond to sampling points along the probe trajectory.
Picture of someone holding the probe that measures the flow field behind an airplane in the wind tunnel.
Andreas Müller, Andrin Landolt and T.K. Rösgen (2012). Probe capture for quantitative flow visualization in large scale wind tunnels. 28th Aerodynamic Measurement Technology, Ground Testing, and Flight Testing Conference
Andreas Müller (2017). Real-Time 3D Flow Visualization Technique with Large Scale Capability. PhD thesis, ETHZ collection.
Andreas Müller (2017). Demonstration of a Real-Time 3D Flow Visualization Technique with Large Scale Capability. 52nd Annual SATA Conference
A. Gisberts, G. Metta (2013).Real-time model learning using Incremental Sparse Spectrum Gaussian Process Regression. Neural Networks, vol. 41, pp. 59-69.