Seminar "Multi-Agent Deep Reinforcement Learning for Distributed Control of Wall-Bounded Turbulent Flows" by Giorgio Cavallazzi

[Speaker]

Giorgio Cavallazzi
Research Student
Department of Engineering
City St George's, University of London

[Abstract]

Wall-bounded turbulent flows exhibit complex regeneration mechanisms through the wall cycle, where turbulent structures near the wall surface drive the perpetuation of turbulence. Traditional flow control approaches typically employ uniform actuation strategies across the entire wall surface, which may not optimally address the inherently heterogeneous nature of turbulent structures. In our research, we present a novel computational approach that introduces spatially distributed control by segmenting the wall into spanwise-oriented strips, each governed by an independent deep reinforcement learning agent.

Our methodology integrates PyTorch-based reinforcement learning with direct numerical simulation using a finite differences solver, creating a multi-agent framework where each agent independently learns optimal policies for local boundary condition modification. The implementation leverages MPI-based parallel computing to manage the computational demands of concurrent agent training while maintaining scalability for high-fidelity turbulence simulations.

Preliminary findings indicate that this distributed control strategy outperforms uniform actuation methods in drag reduction effectiveness, with agents exhibiting emergent coordination behavior through their mutual influence on the shared flow field. The system demonstrates particular success in preserving coherent near-wall velocity streaks while dynamically responding to local flow variations. Analysis of the learned policies and training procedures provides valuable insights into fundamental near-wall turbulence physics and demonstrates the broader potential of multi-agent reinforcement learning in computational fluid dynamics, opening new avenues for sophisticated flow control methodologies with applications in aerospace, automotive, and energy systems.