Embedded Files
The next seminar is on April 10, 2026!
Remote Colloquium on Vortex Dominated Flows (ReCoVor) is an online seminar series that emerged out of the need to facilitate scientific engagement in the face of the COVID-19 pandemic. Widespread social-distancing measures had handicapped what had historically been a fundamental tenet of scientific inquiry - the exchange of new ideas, critical feedback, and engagement with the broader scientific community. In view of this challenge, ReCoVor was created to serve as a forum for encouraging scientific discussion with a focus on graduate students and early stage researchers. ReCoVor was also meant to provide a platform for these researchers to regain some of the opportunities lost for presenting their work to a larger scientific community and for networking, which had resulted from cancelled conferences, collaborative visits, on-campus seminars, etc. Despite the fact that the pandemic is now in our rear-view mirror, there has been overwhelming support for continuing this online series, and, in fact, the membership and participation in the series have continued to grow.
As suggested by the name, this colloquial series is focused on the flow physics of unsteady, vortex-dominated flows, particularly as it applies to fluid-structure interaction, bioflight/swimming, physiological flows, massively separated flows, and other such shear flows. If the flow is unsteady and it involves multiple interacting vortices that induce important effects on the flow, then this research probably belongs in this colloquium. Experimental, computational, and/or analytical contributions are all welcome.
Rajat Mittal (JHU), Jeff Eldredge (UCLA), Anya Jones (UCLA), Karen Mulleners (EPFL), Karthik Menon (Georgia Tech)
Diederik Beckers (Caltech) & Hanieh Mousavi (UCLA)
Upcoming seminar: April 10, 2026
Upcoming seminar: April 10, 2026
106A: Decoding Convolutional Feature Learning in Flapping-Wing Flows
106A: Decoding Convolutional Feature Learning in Flapping-Wing Flows
Arvind Arasu, Indian Institute of Technology, Madras
(student talk)
PI: Sunetra Sarkar
Abstract: Convolutional neural networks (CNNs) are increasingly used to model fluid flows, yet studies on how to design these networks for unsteady flows with moving boundaries—such as flapping wings—remain scarce. In practice, architecture choices have been guided largely by intuition and empirical heuristics rather than by a clear understanding of how these networks process flow data. This limitation becomes especially important in moving-boundary problems, where time-varying discontinuities and complex vortex interactions make effective network design far from straightforward. To address this, we use interpretability techniques to uncover how CNNs learn from flow data, and use those insights to guide principled model design and optimization. We introduce a new saliency-mapping method based on the Convolutional Neural Feature Matrix (CNFM), which provides a layer-by-layer view of what a CNN learns from flow fields. The method not only shows which architectural choices perform well, but explains why they work, enabling targeted redesign instead of blind trial and error. We demonstrate this approach on two tasks: aerodynamic load estimation from instantaneous flow fields and flow reconstruction with an autoencoder, for a pitching-plunging airfoil at Re=300, across three phase offsets spanning distinct wake topologies in both periodic and quasi-periodic regimes. The results make the case for interpretability-guided design. With CNFM-guided redesign, drag coefficient prediction errors fall from 28% to around 2% on unseen quasi-periodic cases; autoencoder reconstruction errors drop from 60% to approximately 5%. Overall, this work demonstrates that interpretability can provide principled guidance for designing efficient, explainable neural networks, with broader applicability beyond CNNs to architectures such as transformers.
106B: TBA
106B: TBA
Details
Details
Format
Format
Two 15-minute talks, each followed by 10 minutes of Q&A
Two 15-minute talks, each followed by 10 minutes of Q&A
When?
When?
Friday, every two weeks, at 11am Eastern Time / 8am Pacific Time.
Friday, every two weeks, at 11am Eastern Time / 8am Pacific Time.
Where?
Where?
Zoom, of course! A Zoom link with password will be provided to registered members.
Zoom, of course! A Zoom link with password will be provided to registered members.
Contact
Contact
Page updated
Report abuse