Скачать с ютуб PhD defense - Daniele NOTO в хорошем качестве

PhD defense - Daniele NOTO

Abstract : In this thesis, we explored the potential for improving Machine Learning models by integrating physical knowledge of dynamic systems. Our first focus was the Lorenz model, a well-established benchmark for chaotic dynamic systems. By applying "curriculum learning" techniques, specifically using entropy to select the most informative regions of the phase space, we achieved notable improvements in the learning process. We then extended our investigation to the domain of multi-scale dynamic systems, particularly shell models of turbulence. We developed an innovative training algorithm that leverages temporal correlation analysis and the inherent dynamics of recurrent neural networks. This approach showed significant advancements in understanding and predicting the behaviour of turbulent systems. Additionally, we sought improvements in reduced-order models (ROM) using Machine Learning. Inspired by existing models, we applied this methodology to the Sabra shell model for turbulence. Our Machine Learning model successfully reproduced missing statistical information, compensating for gaps in the datasets. Looking ahead, we propose that the CD-ROM approach could be extended to address more complex problems, such as 2D or 3D turbulence, or to analyze real data from direct numerical simulations of the Navier-Stokes equations. This extension holds promise for broadening the application of Machine Learning to the study of complex dynamic systems, which could provide significant insights and advancements.