Conditional Neural Field Latent Diffusion Model for generating Spatiotemporal Turbulence (Pan Du)

Описание: We were fortunate to hear about a generative model framework for complex turbulent flow simulations that can support a number of useful tasks, ranging from uncertainty quantification to flow super-resolution to reconstruction from sparse sensor measurements. This was work presented by Pan Du, currently a Ph.D. candidate at the University of Notre Dame. More details can be found below, and in his related paper: https://www.nature.com/articles/s4146....

Check out and enjoy this great talk with some beautiful visualizations of turbulence, and consider subscribing to this channel to be notified of future seminars.

Abstract:
Pan will present the CoNFiLD model, a novel generative framework for simulating complex turbulent flows in 3D irregular domains. While traditional eddy-resolved simulations are accurate, their high computational cost limits usability. CoNFiLD addresses this by integrating neural field encoding with latent diffusion, enabling efficient, probabilistic modeling of spatiotemporal dynamics. It supports a wide range of tasks—such as flow super-resolution, sparse reconstruction, and data restoration—via Bayesian conditional sampling, all without retraining. Results across diverse turbulent scenarios highlight its potential for advancing data-driven turbulence modeling.

Bio: Pan Du received his bachelor's degree in Thermal Engineering from Tsinghua University and completed his master's in Mechanical Engineering at Washington University in St. Louis. He is currently a Ph.D. candidate in Aerospace and Mechanical Engineering at the University of Notre Dame under the guidance of Prof. Jian-Xun Wang. Pan's research spans multiple disciplines, including scientific machine learning, Bayesian inference, uncertainty quantification, geometric deep learning, and computational fluid mechanics.

00:00 Start
02:43 Data-driven Turbulence Modeling
10:40 Generative AI Techniques
13:37 CoNFiLD and its Applications
20:21 Computational Cost of the Generative Framework
22:45 Zero-shot Conditional Generation
28:50 Super-resolution
30:46 Conclusion, Useful Links and Q&A