https://www.linkedin.com/in/matteorosellini/
Contacts:
Ciclo: XXXIX
Curriculum: Aerospace Engineering
Title of the PhD Project: Enhancing turbulence modeling with Machine Learning
Supervisor(s): Prof.ssa Maria Vittoria Salvetti, Prof. Alessandro Mariotti, Dott. Oriol Lehmkuhl
In collaboration with: Numerical Mathematics Lab in Biorobotics Institute, Scuola Sant’Anna; Large Scale Computational Fluid Dynamics Group in Barcelona Supercomputing Center (BSC).
Abstract of the PhD project:
Recent progress in computational sciences expands the possibilities of applying Machine Learning to engineering problems. Fluid mechanics is characterized by the management of large datasets obtained from experiments, field measurements, and numerical simulations. In this context, Machine Learning provides a framework capable of addressing major challenges, with turbulence modeling representing a primary target. The project is guided by the objective of constructing a framework that is physics-based, interpretable, and generalizable, with applications both to RANS simulations under the Boussinesq assumption and to wall modeling in LES.
The research focuses on the development of turbulence models for RANS simulations, enhanced through data assimilation techniques followed by regression phases based on Machine Learning. In parallel, wall-modeling strategies for LES are investigated, starting from equilibrium formulations and progressively incorporating non-equilibrium effects. To rigorously address these modeling challenges, uncertainty quantification tools are employed to perform sensitivity analyses of the governing parameters. A significant part of the work is dedicated to exploring and implementing methodologies in uncertainty quantification.
Studies on RANS closures are applied to flows around rectangular cylinders with different aspect ratios and Reynolds numbers. The baseline investigations are carried out in two dimensions, with RANS equations solved in FreeFem++ and regression implemented in Python. Sensitivity analyses are conducted on the flow features selected for training. The methodology is then extended to three-dimensional urban-like configurations, ranging from a single block to arrays of blocks, with the long-term objective of simulating realistic portions of a city. These test cases include geometries with available experimental data, and the simulations account for ground effects and a developed atmospheric boundary layer.
For uncertainty quantification, surrogate models are adopted to reduce the computational cost of high-dimensional problems. Polynomial Chaos Expansion is applied for one or two uncertain parameters, while sparse grid techniques are employed for higher-dimensional cases. To improve efficiency, an adaptive sparsification strategy based on a zero-cost index is proposed and validated on both analytical functions and a real-world case.
In the LES framework, wall-modeling approaches are explored from equilibrium-based formulations to those incorporating non-equilibrium effects. Sensitivity analyses are performed on key parameters, particularly the exchange height separating the resolved region from the modeled near-wall layer. A novel damping function for turbulent viscosity is introduced to suppress kinetic energy production during relaminarization, showing promising results in recirculating regions.
Finally, contributions are made to the ROSAS TC05 project at the University of Pisa, including high-fidelity WRLES simulations and experimental setups for PIV-based velocity measurements. These efforts aim to provide a robust numerical–experimental comparison for flow over a rounded ramp, supporting broader objectives in aeronautical design and climate-neutral innovation.
Workshops:
• Von Karman Institute for Fluid Dynamics and Universitée Libre de Bruxelles, Machine Learning for Fluid Mechanics: Analysis, Modeling, Control and Closures, Brussels, 2024 January 29th – February 2nd.
• ERCOFTAC, Workshop on Machine Learning for Fluid Dynamics, 6th - 8th March 2024, Sorbonne University, Paris, France
Scientific Contribution:
• Co-author: Near-the-wall analysis of the NASA Wall-Mounted Hump APS-DFD, Houston (Texas, US), 23–25 November 2025
• Oral presentation: Assessing the Impact of Exchange Location Height on High-Order Wall Modeled LES ERCOFTAC ETMM-15, Dubrovnik (Croatia), 21–25 September 2025
• Co-author: A Surrogate-Informed Sparse-Grid Approach for Flashback Prediction in H2-Fueled Perforated Burners ERCOFTAC ETMM-15, Dubrovnik (Croatia), 21–25 September 2025
• Co-author: Data Augmented RANS Modeling for Rectangular Cylinders 7th International Conference on Turbulence and Interactions, La Palma (Spain), 30 March – 4 April 2025
• Oral presentation: Data Augmented RANS Modeling for Rectangular Cylinders EuroMech Colloquium on Data-Driven Fluid Dynamics, London (UK), 2–4 April 2024
• Poster presentation: Enhancing RANS Turbulence Modeling for Rectangular Cylinders using Data Assimilation and Machine Learning ConScience: Towards a Responsible Use of Machine Learning, Pisa (Italy), 17 December 2024
Papers in preparation:
• M. Rosellini, A. Mariotti, G. Stabile, M. V. Salvetti, Enhancing RANS Turbulence modelling for rectangular cylinders using Data Assimilation and Machine Learning
• M. Rosellini, F. Fruzza, A. Mariotti, M. V. Salvetti, L. Tamellini, A Surrogate – Informed Framework for Sparse Grid Interpolation