Categoria:
Seminari di Modellistica Differenziale Numerica
Data e ora inizio evento:
Data e ora fine evento:
Aula:
Altro (Aula esterna al Dipartimento)
Sede:
IAC-CNR, via dei Taurini 19
Aula esterna:
116
Speaker:
Niccolò Tonicello (SISSA)
Turbulence modeling has long been one of the most challenging areas in computational fluid dynamics. At the same time, turbulence is also one of the most ubiquitous phenomena in fluid dynamics, appearing in contexts ranging from environmental atmospheric flows to jet spray atomization in combustion engines. Despite decades of progress in high-performance computing, physical modeling, and efficient numerical methods, the application of data-driven strategies to turbulence has only emerged in the past few years. The interest in machine-learning-based approaches for turbulence modeling has been growing rapidly. In this talk, I will present two main lines of research that we are pursuing in data-driven turbulence modeling. The first focuses on nonlinear reduced-order models for chaotic systems. In particular, we have recently explored the use of specialized AutoEncoders, called symmetric AutoEncoders, which appear particularly well-suited for capturing the complex dynamics often encountered in turbulent flows. The second part of the presentation will focus on a fully end-to-end differentiable solver, enabling model training directly within the simulation loop, with a particular emphasis on sub-grid scale modeling in Large-Eddy Simulations.

