Bulletin of the American Physical Society
72nd Annual Meeting of the APS Division of Fluid Dynamics
Volume 64, Number 13
Saturday–Tuesday, November 23–26, 2019; Seattle, Washington
Session P20: Experimental Techniques: Data Analysis, Bias and Uncertainty |
Hide Abstracts |
Chair: Barton Smith, Utah State University Room: 602 |
Monday, November 25, 2019 5:16PM - 5:29PM |
P20.00001: Inferring Physical Laws from Data: Disambiguating Gravity from Fluid Forces on Falling Objects Brian De Silva, David Higdon, Steven Brunton, Nathan Kutz Machine learning and artificial intelligence algorithms are now being used to automate the discovery of governing physical equations from measurement data alone. However, positing a universal physical law from data is challenging without simultaneously proposing an accompanying discrepancy model to account for the inevitable mismatch between theory and measurements. By revisiting the classic problem of modeling falling objects of different size and mass, we highlight a number of subtle issues that must be addressed by modern data-driven methods for the automated discovery of physics. Specifically, we show that measurement noise and complex secondary physical mechanisms, such as unsteady fluid drag forces, can obscure the underlying law of gravitation, leading to a faulty model. Using the sparse identification of nonlinear dynamics (SINDy) algorithm, with the added assumption that each falling object is governed by the same physical law, we are able to identify a viable discrepancy model to account for the fluid dynamic forces that explain the mismatch between a posited universal law of gravity and the measurement data. This work highlights the fact that the naive application of ML/AI will generally be insufficient to extract universal physical laws without further modification. [Preview Abstract] |
Monday, November 25, 2019 5:29PM - 5:42PM |
P20.00002: Data Analysis of Hybrid Rocket Fuels Combustion Tests Alexander Ruettgers, Anna Petrarolo, Mario Kobald Clustering techniques were applied to hybrid rocket combustion tests to better understand the complex flow phenomena. Novel techniques such as hybrid rockets that allow for cost reductions of space transport vehicles are of high importance in space flight. However, the combustion process in hybrid rocket engines is still a matter of ongoing research and not fully understood yet. Recently, combustion tests with different paraffin-based fuels have been performed at the German Aerospace Center (DLR). For a better understanding of the experiments, the combustion process has been captured with a high-speed video camera, which leads to a huge amount of images for each test. In order to catch the essential flow structures, the combustion dataset has been analyzed with unsupervised machine learning techniques. In this talk, we present the outcome of the clustering. Using machine learning techniques, valuable insights into the different combustion phases were obtained and a comparison of the quality of the combustion flame in the different tests could be made. In particular, depending on the fuel formulation and oxidizer mass flow, differences in the transients and flame brightness were found. [Preview Abstract] |
Monday, November 25, 2019 5:42PM - 5:55PM |
P20.00003: Data assimilation method to de-noise and de-filter particle image velocimetry data Roland Bouffanais, Jurriaan J. J. Gillissen, Dick K. P. Yue We present a variational data assimilation method in order to improve the accuracy of velocity fields $\tilde{\mathbf{v}}$, that are measured using particle image velocimetry (PIV). The method minimizes the space-time integral of the difference between the reconstruction $\mathbf{u}$ and $\tilde{\mathbf{v}}$, under the constraint, that $\mathbf{u}$ satisfies conservation of mass and momentum. We apply the method to synthetic velocimetry data, in a two-dimensional turbulent flow, where realistic PIV noise is generated by computationally mimicking the PIV measurement process. The method performs optimally when the assimilation integration time is of the order of the flow correlation time. We interpret these results by comparing them to one-dimensional diffusion and advection problems, for which we derive analytical expressions for the reconstruction error. [Preview Abstract] |
Monday, November 25, 2019 5:55PM - 6:08PM |
P20.00004: State-space Optimized Dynamic Mode Decomposition for Noisy Data Taku Nonomura, Kazuyuki Nakamura, Naoto Nakano, Steven L. Bruntun, J. Nathan Kutz This presentation proposes several new formulations of dynamic mode decomposition (DMD) for full-state measurements of a linear dynamical system with process and measurement noise. First, we develop two methods to denoise and reconstruct the true state of the approximated linear system from noisy experimental data: the DMD-based state variable reconstruction (DMDsvr) and the DMD-based state space reconstruction (DMDssr). DMDsvr estimates the state variable as a solution of a least square problem, when the system coefficients and the noise variances are known. DMDssr simultaneously estimates the noise variance and the state variables by the expectationâ€“maximization (EM) algorithm in a Bayesian framework. The final method, state-space optimized DMD (ssoDMD), simultaneously estimates the DMD coefficients together with the noise variances and the state variable. The proposed ssoDMD can estimate the system coefficients, noise variance and true state variables from noisy data. Estimation of system coefficients and noise variances can be used for data-assimilation using Kalman filter. Numerical tests show an improvement of the proposed methods over conventional DMD for linear systems with process noise. [Preview Abstract] |
Monday, November 25, 2019 6:08PM - 6:21PM |
P20.00005: 4D Flow MRI Bias Error Estimation Sean Rothenberger, Jiacheng Zhang, Melissa Brindise, Susanne Schnell, Pavlos Vlachos, Vitaliy Rayz 4D flow MRI is a non-invasive imaging technique providing time-resolved, volumetric, 3-directional velocity information of a patient's cardiovascular flow in vivo. This technique suffers from low VNR and resolution. `Enhanced' 4D flow MRI measurements can be produced by informing the measurements with high-resolution flow modeling methods, e.g. CFD, PIV. However, the error of the 4D flow MRI measurements must first be defined. Past literature defines a model for measurement error, but neglects bias error. This study estimates bias error by integrating an approximated intravoxel velocity profile within the limits of the voxel dimensions. The model of bias error was tested in synthetic flow with a velocity range of 2m/s. The effect of the voxel size and noise was determined by investigating a range of voxel sizes and noise standard deviations of 1 to 0.147mm$^{\mathrm{3}}$ and 0 to 6cm/s, respectively. The estimates were compared to the true bias error calculated using the true velocity profile. Results show that the bias error is proportional to the square of the voxel size. The estimated bias error was determined to be fourth-order accurate when the RMS difference was compared to the change in voxel size. The bias error estimate is proportional to the square of the noise's standard deviation. [Preview Abstract] |
Monday, November 25, 2019 6:21PM - 6:34PM |
P20.00006: Unscented Kalman filter (UKF) based nonlinear parameter estimation for a turbulent boundary layer:a data assimilation framework Zhao Pan, Yang Zhang, Jonas Gustavsson, Jean-Pierre Hickey, Louis Cattafesta A turbulent boundary layer is an essential flow case of fundamental and applied fluid mechanics. However, accurate measurements of turbulent boundary layer parameters (e.g., friction velocity $u_\tau$ and wall shear $\tau_w$), are challenging, especially for high speed flows. Many direct and/or indirect diagnostic techniques have been developed to measure wall shear stress. However, based on different principles, these techniques usually give different results with different uncertainties. The current study introduces a nonlinear data assimilation framework based on the Unscented Kalman Filter that can fuse information from i) noisy and gappy measurements from Stereo Particle Image Velocimetry, a Preston tube, and a MEMS shear stress sensor, as well as ii) the uncertainties of the measurements to estimate the parameters of a turbulent boundary layer. A direct numerical simulation of a fully developed turbulent boundary layer flow at Mach 0.3 is used first to validate the data assimilation algorithm. The algorithm is then applied to experimental data of a flow at Mach 0.3, which are obtained in a blowdown wind tunnel facility. The UKF-based data assimilation algorithm is robust to uncertain and gappy experimental data and is abl [Preview Abstract] |
Monday, November 25, 2019 6:34PM - 6:47PM |
P20.00007: Power Law Decay Estimation for Turbulent Spectral Densities Carl R. Hart, Gregory W. Lyons, Nathan E. Murray Turbulent flows commonly feature power law decay in one or more field quantities, such as the -5/3 inertial subrange power law for velocity spectra. Assuming sufficient time series data are collected, the problem of estimating the power law decay rate of a turbulent spectral density relies on two factors: the correct choice of data window in statistical signal processing, and an objective procedure to estimate the power law decay rate. In this context the single most important factor for a data window is the side-lobe decay rate. Ensuring the side-lobe decay rate exceeds that of measured data avoids the subtle error of spectral leakage. An objective procedure to estimate the power law decay rate is based on a maximum likelihood estimator. Under the assumption that the Fourier transform of turbulence time series is a circularly-symmetric complex normal random variable a likelihood function for the power spectral density is based on a Gamma distribution. Maximizing the log-likelihood, with a spectral model that parameterizes the Gamma distributions, leads to a robust estimator for the power law decay rate. These concepts are illustrated through synthetic realizations of colored noise, acoustic measurements of a supersonic turbulent jet, and atmospheric surface-layer turbulence. [Preview Abstract] |
Monday, November 25, 2019 6:47PM - 7:00PM |
P20.00008: Identification of Dynamic Atmospheric Conditions via Total Variation Nicholas Hamilton Selection of atmospheric events that conform to particular conditions of interest within multivariate data is necessary to validate of emerging high-fidelity simulations of wind plant flows. Conditions of interest are frequently determined simply as those that occur most often, given the need for well-converged statistics from observations. Aggregation of observations without regard to covariance between time series discounts the dynamical nature of the atmosphere and is not sufficiently representative of wind plant operating conditions. Identification and characterization of continuous time periods representative of atmospheric conditions that have a high value for analysis or simulation sets the stage for validation of more advanced physical mechanisms. The total variation of the atmosphere is a metric that takes into account variability within each channel as well as covariance between channels and identifies periods of interest that conform to desired objective functions, such as quiescent conditions, wind speed ramps or waves, or sudden changes in wind direction. Direct identification and classification of events of interest within atmospheric data sets is vital to developing our understanding of wind plant response and to the formulation of forecasting and control models. [Preview Abstract] |
Monday, November 25, 2019 7:00PM - 7:13PM |
P20.00009: Using Machine Learning to Determine the Velocity Information Content in OH-PLIF Images Shivam Barwey, Malik Hassanaly, Venkat Raman, Adam Steinberg This study determines the velocity field information contained purely in OH-PLIF images in the closed domain of a premixed swirl combustor. A fully convolutional neural network (CNN) is used with a dataset containing simultaneous OH-PLIF and PIV measurements in both attached and detached flame regimes. To facilitate the study, the CNN represents a direct projection from OH-PLIF to PIV field. Two types of models are trained: 1) a global CNN which is trained using images from the entire domain, and 2) a set of local CNNs which are trained only on individual sections of the domain. Local models show improvement in creating PIV fields in both attached and detached regimes over the global models in most settings. A comparison between model performance in attached and detached regimes shows that the CNNs are much more accurate across the board in creating velocity fields for attached flames. Further, time history inclusion in the OH fields is also studied, as is the ability of the model to extrapolate to unexplored regions of the domain. Ultimately, this work shows that there is redundant information in the OH-PLIF images, and can open the door for the development of diagnostic tools that decrease the overlapping content between simultaneously measured fields. [Preview Abstract] |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2020 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
1 Research Road, Ridge, NY 11961-2701
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700