From Energy Spectra to Velocity Fields: Machine Learning for Turbulent Flow Reconstruction

with Anne Staples,
Associate Professor, Virginia Tech Mechanical Engineering

Friday, March 7, 10:10 am
440 Goodwin Hall, Blacksburg
or room 6-051, VTRC, Arlington 

In this talk, I will present preliminary results from a novel application of machine learning for 2D turbulent vorticity field reconstruction. Recently, my colleagues and I introduced an encoder-decoder Y-network model designed as a lifting operator to reconstruct isotropic decaying turbulent vorticity fields from their instantaneous energy spectra. To capture different turbulence features, we trained two specialized models targeting small-scale and large-scale structures, respectively. These models were then combined through a transfer learning framework, leveraging their encoders and integrating their outputs via a new decoder in the Y-network architecture. This innovative approach led to significant performance improvements, reducing mean squared error (MSE) and mean absolute error (MAE) by 65% and 40%, respectively, for the reconstruction task.

I will also introduce preliminary results using generative AI, specifically a conditional diffusion model, as an alternative lifting operator, which appears to be highly generalizable for reconstructing turbulent flow fields.

Finally, I will present our prior work using an encoder-decoder model with stacked LSTM layers to map energy spectra to velocity fields in a 1D Burgers' turbulence model. By integrating this model into a coarse projective integration multiscale simulation framework, we achieved a 443-fold acceleration in the flow's evolution to statistical stationarity (Dhingra et al., Phys. Fluids, 2024). Preliminary results suggest that the current Y-network and conditional diffusion models can similarly be integrated into multiscale simulations, enabling faster and more accurate simulations of turbulent flows, while preserving the full spatial resolution of the flow field.

Anne Staples is an Associate Professor of Mechanical Engineering at Virginia Tech, where she directs the Laboratory for Fluid Dynamics in Nature (FINLAB). She earned her Ph.D. in Mechanical and Aerospace Engineering from Princeton University and completed a National Academies postdoctoral fellowship at the Naval Research Laboratory.

Her research focuses on fundamental flow physics, computational fluid dynamics (CFD) algorithms, and fluid dynamics in biological, natural, and medical contexts. Her group investigates topics such as insect respiration, coral hydrodynamics, and drug-delivery systems using computational modeling and microfluidic experiments. FINLAB has pioneered insights into valveless flow control in insect-inspired systems, and Dr. Staples has been featured on NPR’s All Things Considered as an expert in insect fluid dynamics.

Dr. Staples has received numerous honors, including the Virginia Tech Engineering Scholar of the Week, the Dean’s Award for Excellence in Teaching, and a Fulbright U.S. Scholar Award. In 2024, she was awarded the NIH NIBIB Trailblazer R21 Award, recognizing her early-stage, high-impact research at the interface of engineering and the life sciences.