Generative AI for particle trajectories in turbulence: innovative study by the Statistical Mechanics and Complex Systems Physics Group
The prestigious journal Nature Machine Intelligence has published research conceived and carried out by researchers in our department, placing it on the cover for its April 2024 issue.
The statistical mechanics group led by Luca Biferale, full professor of Theoretical Physics at this Department, has developed a generative model using advanced Machine Learning techniques based on diffusion algorithms (data-driven diffusion model) capable of generating synthetic trajectories of particles in turbulent flows, the focus of many applied and fundamental problems related to chaotic dispersion in engineering, bio-fluidics, atmospheric physics, oceanography, and astrophysics.
Despite outstanding theoretical, numerical, and experimental efforts over the past 30 years, no existing model had been able to faithfully reproduce the statistical and topological properties exhibited by particle trajectories in turbulence.
Through careful statistical analysis, they were able to show that machine learning models can "capture" the full complexity of turbulent dynamics and generalize to predict extreme and rare events not observed during training, demonstrating how these new data-driven techniques are capable of going beyond the generation of "synthetic" text or images by opening new avenues to increase the data available to scientists even in "quantitative" fields, as in the case of their research on complex fluids and flows.
This is an important first result of a long-term project in which the group of young researchers formed by Michele Buzzicotti (RTDb), Tianyi Li (post-doc), Fabio Bonaccorso (technologist) and Martino Scarpolini (post-doc) have been working on advanced problems at the frontier between artificial intelligence and theoretical physics, together with post-docs Robin Heinonen, Fabio Guglietta and Lorenzo Piro and PhD students Chiara Calascibetta, Damiano Capocci, Andre Freitas, Francesco Fossella and Elisa Bellantoni, funded at European level by the European Research Council and at national level by the Ministry of Research through several projects: ERC Advanced Grant, Aqtivate, MUR-FARE, PRIN2022.
