Abstract:

Lagrangian turbulence lies at the core of numerous applied and fundamental problems. However, despite decades of theoretical, numerical, and experimental research, no existing model can accurately reproduce particle trajectories’ statistical and topological properties in turbulent flows. This talk presents a machine learning framework based on a state-of-the-art diffusion model to generate single particle trajectories in three-dimensional turbulence at high Reynolds numbers [1]. This approach bypasses the need for direct numerical simulations or experiments to obtain reliable Lagrangian data. Our results show that the model reproduces key statistical features across time scales, including fat-tailed velocity increment distributions, and anomalous scaling laws. Additionally, we extend this method to reconstruct missing spatial and velocity data along trajectories of small objects passively advected by turbulent flows, such as oceanic drifters from NOAA’s Global Drifter Program [2]. The method accurately reconstructs velocity signals while preserving non-Gaussian, intermittent scale-by-scale properties. Notably, the model is flexible enough to handle different data gap configurations and to exploit correlations enabling superior performance over traditional Gaussian Process Regression methods.

Introduction of speaker:

Dr. Michele Buzzicotti received his Ph.D. in Physics from the University of Rome Tor Vergata in 2017. He is currently a tenure-track researcher in the Department of Physics at the same university. Dr. Buzzicotti has held visiting research appointments at the Eindhoven University of Technology and the University of Rochester in New York.  His research lies at the intersection of turbulence, complex fluid dynamics, and data-driven modeling, with a strong focus on integrating machine learning and artificial intelligence in studying multiscale physical systems. He is the Principal Investigator of the national project, "Data-driven and equation-based tools for deep understanding of multi-scale complex turbulent flows" (2025-2028), which is funded by the FIS2 initiative for excellence in research. He has published over 40 peer-reviewed papers, which have been cited more than 1100 times, giving him an h-index of 20. His work has appeared in prestigious journals such as Nature Machine Intelligence, Nature Communications, Physical Review Letters, and the Journal of Fluid Mechanics. Dr. Buzzicotti has played an active role in major European and international research consortia, including the ERC Advanced Grant (AdG) projects Smart-TURB and NewTURB, which were led by Prof. Luca Biferale. He has also contributed to several PRACE and EuroHPC high-performance computing projects. In addition to his research, Dr. Buzzicotti has taught courses in statistical mechanics, machine learning for physics, and turbulence and complex fluids at both the undergraduate and graduate levels. He is a review editor for Frontiers in Physics and a guest editor for the European Physical Journal E. He is also a frequent reviewer for high-impact journals, including PRL, Nature Communications, and JFM. Dr. Buzzicotti is an active member of the INFN, EUROMECH, and the APS. He collaborates with leading scientists worldwide on developing physics-informed, data-driven models for complex flows.

Dr. Michele Buzzicotti is a tenure-track researcher in Physics at the University of Rome Tor Vergata. His work focuses on turbulence and complex fluids, leveraging machine learning for data-driven modeling of multiscale systems. He is the Principal Investigator of a national FIS2 project (2025-2028) and has authored over 40 publications (h-index 20) in prestigious journals including Nature Machine Intelligence and Physical Review Letters. He is an active member of INFN, EUROMECH, and APS. His talk will be Reconstruction of Lagrangian Turbulent data with data-driven Generative Models. I will give the time to him. Let’s welcome Michele.

审核：刘有晟、游小清