Announcement Detail

Tuesday, September 23, 2025

2:00 PM EDT

Join via Zoom: https://us06web.zoom.us/j/82464478256?pwd=ZMkJVFdjMJzadgnVWFPqsdUSs4qTaY.1
 

Student Chapter Seminar Series

Joint space-time wind field data extrapolation using sparse representations and nonparametric Bayesian dictionary learning

Speaker

George Pasparakis, Johns Hopkins University

Abstract:

Accurate reconstruction and extrapolation of wind field data and related statistics is critical in assessing wind-induced structural response. Nevertheless, both from a simulation and a structural health monitoring perspective various constraints, such as computational memory overheads and sensor network limitations, often dictate the reconstruction of wind field data from a limited number of observations. This talk presents a unified framework for wind field extrapolation and stochastic field statistics estimation, developed through sparse representations and dictionary learning techniques. Specifically, compressive sampling methodologies are introduced, and the application of ℓp-norm minimization techniques is showcased within the context of one-dimensional wind data reconstruction. Further, nuclear norm minimization techniques are discussed as a computationally efficient alternative for addressing memory limitations associated with higher-dimensional domains. These approaches enable the recovery of incomplete time-histories and the estimation of power spectral densities, cross-correlations, and coherence functions, even under considerable degrees of missing data. Building upon these concepts, an enhanced nonparametric Bayesian dictionary learning framework is introduced, which is advantageous with respect to standard compressive sampling strategies. The proposed scheme adaptively identifies low-dimensional representations directly from data, circumvents the need for an a priori basis construction and quantifies the uncertainty in the estimates. The efficacy of the proposed frameworks is illustrated through simulated data generated via the stochastic wave spectral representation method, based on a prescribed joint wavenumber-frequency spectrum, and boundary layer wind tunnel measurements which exhibit strong spatial variability and non-Gaussian statistics.

Bio

George Pasparakis is a Postdoctoral Research Fellow at the Hopkins Extreme Materials Institute (HEMI) at Johns Hopkins University, working with the Shields Uncertainty Group (SURG) and the Graham-Brady research group. He earned his Ph.D. in Civil Engineering with highest honors from Leibniz University Hannover where his work was supported by the Marie Skłodowska-Curie Actions (MSCA) fellowship. He also holds a Diploma in Mechanical Engineering from the University of Thessaly. His research interests focus on stochastic dynamics, uncertainty quantification, Bayesian neural networks, computational mechanics, and advanced signal processing. He applies these methods to a range of engineering problems, including vibration energy harvesting, wind field reconstruction and extrapolation, and uncertainty quantification in materials modeling under extreme conditions.