VKI Seminar Series 2024

13 June 2024
2:00 PM – 3:00 PM (Brussels Time)

REGISTRATION

Multi-scale chemically reacting flows: model reduction and machine learning (open to public)

Guest Speaker: Prof. Riccardo Malpica Galassi, assistant professor, Sapienza University of Rome

Abstract: High-fidelity simulations of multi-dimensional turbulent reacting flows at large Reynolds and Karlovitz numbers with detailed chemistry and transport are an essential tool to provide physical insights and to serve as a valuable database for engineering model validation. Despite rapid advances in CPU/GPU high performance computing hardware, the challenge remains due to the large number of reactive scalar variables and the wide spectrum of spatial and temporal scales. Therefore, there are continuing motivations to develop advanced reduced order models (ROM) to improve computational efficiency while preserving the fidelity and predictive capability of the simulations.
In describing the temporal evolution of chemically reacting systems, the large number of characteristic chemical time scales associated with individual reaction pathways cause the stiffness problem, often demanding an unnecessarily large number of time steps to integrate the equations to the desired practical time.  In fact, the fastest chemical scales are orders of magnitude smaller than the flow scales of interest.
Model reduction takes effect by recognizing that fast processes constrain the slow dynamics to evolve on a lower-dimensional, invariant manifold. To this end, the computational singular perturbation (CSP) framework employs an eigenvalue decomposition of the local Jacobian matrix to identify the fast/slow spectral gap and projects the system onto the slow invariant manifold (SIM), allowing a significantly accelerated time integration with an efficient explicit algorithm, along with a corrective projection for fast exhausted modes. In fact, the slow dynamics, free of the fast scales, is not stiff anymore and evolves within the SIM at a pace which is orders of magnitude larger than the system's fastest timescale. While the CSP-based solvers have demonstrated effective computational accelerations by orders of magnitude in time steps, a major computational overhead remains in the operation of the large Jacobian matrix to compute the local CSP projection basis. Recently, the renewed interest in CSP-based solvers has been catalyzed by the advent of machine learning techniques. Data-driven approaches may be fruitfully employed to learn projection operators and non-stiff latent spaces, enhancing the efficacy of CSP in capturing the slow system dynamics and reducing computational complexities.

Biography: Riccardo Malpica Galassi is an assistant professor at Sapienza University of Rome. He received his PhD from the Mechanical and Aerospace Department at the University of Rome in 2018, where subsequently, he spent over three years as a post-doctoral fellow. Following this, he spent two years as a Marie Curie post-doctoral Fellow at the Aero-thermo-mechanics department at Université Libre de Bruxelles.He is currently working on the physical understanding of the processes that characterize combustion phenomena, on reduced order models and digital twins, on multi-scale adaptive solvers, on machine learning for combustion, on uncertainty quantification, and reactive flows CFD, with special interest towards turbulent combustion and spray combustion.

Past VKI Seminars Series 2022

 

Past VKI Seminars Series 2023