Announcement Detail

Wednesday, February 4, 2026

1:00 PM CST

Join via Zoom: https://us06web.zoom.us/j/87331120699?pwd=MExkQ085ZHh1aDc3ZnBwQzFlUHI4UT09

Large Scale Structural Systems and Optimal Design Industrial Colloquium

Sparse-grid-accelerated optimized dynamic mode decomposition for higher-dimensional parametric reduced-order modeling

Speaker

Ionut-Gabriel Farcas, Virginia Polytechnic Institute and State University

Abstract

Many problems in science and engineering require running the same simulation repeatedly for different operating conditions, for example, in optimization, uncertainty quantification, or control. These so-called "many-query" tasks are central to emerging digital-twin efforts, but they are often computationally infeasible when each simulation is expensive and depends on variations in many input parameters. In this presentation, we introduce an efficient data-driven approach for building parametric reduced-order models that remain accurate while drastically reducing computational cost. Rather than relying on dense parameter sweeps, which become prohibitively expensive as the number of varying inputs grows, we use a sparse grid strategy to select a small but informative set of parameter instances that capture the dominant trends in the system. This approach mitigates the exponential cost growth typically associated with high-dimensional parameter spaces. We demonstrate the method using gyrokinetic simulations of plasma micro-instabilities relevant to fusion experiments. Reduced-order models are constructed for the full five-dimensional gyrokinetic distribution function by combining optimized dynamic mode decomposition with sparse sampling in parameter space. For a realistic electron-temperature-gradient-driven instability based on a DIII-D discharge involving six input parameters, we show that a predictive parametric reduced-order model can be built using only 28 high-fidelity gyrokinetic simulations. The resulting model is up to three orders of magnitude cheaper to evaluate than full simulations while retaining key physical fidelity. These results highlight the potential of sparse-grid-accelerated, data-driven reduced-order modeling to make high-dimensional many-query tasks tractable, supporting future applications in fusion optimization, uncertainty analysis, and real-time control.

Bio

I am an Assistant Professor in the Department of Mathematics at Virginia Tech. Prior to joining Virginia Tech, I was a Postdoctoral Fellow at the Oden Institute for Computational Engineering and Sciences at The University of Texas at Austin, where I worked with Dr. Karen Willcox and Dr. Frank Jenko on data-driven reduced-order modeling and uncertainty quantification for large-scale simulations in rocket combustion and fusion plasmas. I earned my Ph.D. summa cum laude from the Technical University of Munich in 2020, focusing on efficient numerical methods for uncertainty quantification in computationally intensive problems. My research lies at the intersection of data-driven learning, model reduction, uncertainty quantification, multi-fidelity methods, and high-performance computing, with applications to complex physical systems.