STREAmSML: A GPU-enabled platform for sensing, modeling and control of compressible turbulent flows
Loading...
Authors
Pavlik, Nathaniel
Issue Date
2025
Type
Thesis
Language
en_US
Keywords
boundary layer flow , hypersonic , neural networks , reinforcement learning , shock boundary layer interaction , supersonic
Alternative Title
Abstract
High-speed aerodynamic flows such as supersonic turbulent boundary layers and shock boundarylayer interactions (SBLI) pose significant challenges for flow control due to strong compressibility
effects, shock-induced separation, and nonlinear multiscale turbulence. At the same time, rapid
advances in machine learning have created new opportunities for data-driven sensing, modeling,
and control; provided that standardized, high-fidelity simulation platforms exist to support repro-
ducible evaluation across algorithms. This thesis introduces STREAmSML, a GPU-accelerated,
Gymnasium-compliant platform built upon the validated STREAmS direct numerical simulation
(DNS) solver, enabling systematic investigation of sensing, reduced-order modeling, and closed-loop
actuation for compressible turbulent flows at supersonic Mach numbers.
STREAmSML integrates a modular Python–Fortran/CUDA architecture that permits reinforcement-learning-based, classical, and open-loop jet actuation strategies while preserving the underlying solver’s numerical accuracy and physical fidelity. A span-averaged observation framework reduces the dimensionality of DNS data by two orders of magnitude, retaining the large-scale coherent
structures essential for control while enabling tractable neural-network-based policies.
Using this platform, sparse sensor placement strategies are first developed for both boundary-
layer and SBLI configurations. Interpolatory and greedy algorithms (QR, TPGR) achieve low recon-
struction error with relatively few sensing locations. Reduced-order models based on Dynamic Mode
Decomposition with control (DMDc) are then constructed. For the boundary layer, moderate POD
truncation (90–95% energy) yields superior predictive performance by removing noise-dominated
high-order modes. For SBLI, the flow exhibits a strongly low-rank structure: the first 4–6 energetic
modes capture approximately 90% of the total variance, enabling compact reduced-order models
that isolate dominant separation-bubble and shock-motion dynamics.
Finally, the platform is used to evaluate opposition control and several reinforcement learning
algorithms including DDPG, PPO, and DQN for real-time jet actuation aimed at reducing skin-
friction drag. While the DDPG controller correctly identifies causal relationships between actuation
and downstream wall-shear stress, none of the RL algorithms produce systematic drag reduction
within the explored training horizon, owing to long convective delays, turbulent variability, and
limited sample efficiency. These experiments nevertheless validate the end-to-end interoperability
of STREAmSML with modern RL frameworks and establish a reproducible benchmark for future
algorithmic innovation.
This work provides the first open, GPU-enabled, CTF-inspired platform capable of supporting
sensing, modeling, and control research in supersonic boundary layers and SBLI. Its modular, high-
performance design positions STREAmSML as a foundation for future advances in data-driven flow
control in compressible turbulent regimes.