Neural networks for flow modeling, analysis, and control: bridging the gap between simulations, and experiments

• Grant number: 24/21444-9
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: March 01, 2025
• End date: February 28, 2026
• Field of knowledge: Engineering - Aerospace Engineering - Aerodynamics
• Principal Investigator: William Roberto Wolf
• Grantee: Tarcísio Costa Déda Oliveira
• Host Institution: Faculdade de Engenharia Mecânica (FEM). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated research grant: 13/08293-7 - CCES - Center for Computational Engineering and Sciences, AP.CEPID

Abstract

By leveraging the recent advancements in the application of neural network (NN) models for surrogate modeling of flows -- with successful applications to the analysis of unstable flow problems and active control in computational environments -- we propose the application of such techniques together with new approaches to bring research on the area closer to experimental applications. Real world implementations of flow control systems involve challenges due to diverse phenomena such as unmodeled dynamics, limited sensor data, noisy measurements, response delays, and setups limited by current technology. On the other hand, state-of-art nonlinear flow control techniques are often implemented in ideal computational environments. They include machine learning approaches such as reinforcement learning for training a controller, surrogate modeling for real-time model predictive optimization, as well as surrogate modeling for offline control design. To shorten the gap between simulations and real world applications, we first propose new computational setups where the aforementioned phenomena present in experiments could be taken into account. Furthermore, we aim to include more realistic sensor/actuation setups that could be implemented in real world tests, which is often disregarded in computational studies. The goal of this project is to develop new machine learning approaches to enhance the already developed control techniques on their ability to better deal with adverse phenomena. The proposed implementations include NN-based systems to estimate unmeasured quantities in the flow, feedback of actuator states to take into account non-ideal devices, and the extension of Kalman filters to optimize the nonlinear estimation of states from noisy measurements and imperfect models.