December 16, 2024
From designing safe and efficient aircraft to predicting hurricane paths, many high-stakes applications depend on the simulation of one of nature’s most notoriously unpredictable phenomena: turbulence. For decades, scientists have grappled with formulating computationally affordable yet accurate models that describe how the smallest and largest fluid motions couple together to engender turbulent flows.
Models for turbulence inspired optimism that solutions to seemingly intractable problems were on the horizon. Yet, such models, often referred to as “closures,” are traditionally painstaking to develop and prone to consequential blind spots in that they lack accuracy for flow configurations that are outside of those used for their formulation. However, an ambitious research effort, with the backing of a $2 million grant from the Defense Advanced Research Projects Agency (DARPA), aims to change that by demonstrating how more sophisticated machine learning techniques can create far better models much faster.
“So many applications require accurate models of turbulent flows, which take decades to develop. By the time we find the best closure, the technology it’s made for has moved forward,” explained Fabrizio Bisetti, an associate professor in the Cockrell School of Engineering's Department of Aerospace Engineering and Engineering Mechanics and lead investigator of the research team. “A second challenge in the use of machine learning is that ‘supervised learning’ models being put forth of recent are often narrowly suited for specific applications so that they do not generalize to other flows and, therefore, lack true predictive power.”
Scientists rely on a computational approach called Large-Eddy Simulation (LES) to simulate turbulent flows. It does so by approximating the effect of the small-scale fluid motions with sub-grid scale (SGS) models, that are now being derived from supervised learning algorithms, then allowing for calculation of the whole flow. With this project, titled Lab2Mod, Bisetti’s research group hopes to show that reinforcement learning — machine learning algorithms with impressive abilities to guide their own training and adapt to situations they aren’t explicitly trained for — can replace supervised learning as the state-of-the-art for developing SGS models.
The Work
“We need to show that this method works even when training with noisy experimental data,” Bisetti said. “That’s why gathering a large amount of data from a tunable physical experiment is essential to DARPA’s objective to incorporate predictive simulations in technology development.”
One key disadvantage to supervised learning is that it requires training data from the so-called “ground truth,” which is collected from simulations with higher fidelity models. By contrast, Lab2Mod’s proposed “scientific multi-agent reinforcement learning” (SciMARL) uses agents that interact with their environment, take note of the outcomes and draw inferences about future interactions, all while refining their thought processes and plucking useful information out of noisy experimental data from a real-world flow.
Gathering data on turbulence at conditions relevant to Department of Defense applications typically involves large-scale set-ups like wind tunnels that cost tens or hundreds of thousands of dollars per day to run. Instead, the researchers have opted to generate their flows with a von Kármán Tabletop Turbulence Simulator (VK-TTS), affectionately known as “the French Washing Machine.” The device is compact as it uses two opposing impellers in a small tabletop vessel to create turbulence comparable to that produced in much larger, costlier set-ups. With far greater tunability over a broad range of configurations compared to a wind tunnel, as well as planned upgrades to the diagnostic systems, the VK-TTS offers a wealth of possible configurations and experimental data that should highlight SciMARL’s autonomous learning capabilities. Bisetti and collaborators have experience with the VK-TTS from past research conducted with funding from a National Science Foundation award.
“We envision a situation where the machine learning algorithm will eventually take the lead and adjust the impeller rotation speed to gather informative data at specific conditions,” Bisetti noted.
Why It Matters
If successful, the new framework could illuminate a path through challenges that have at times seemed intractable, specifically the tendency of turbulent flows to undergo large scale changes in their state with important changes to device performance. That would mean better drag reduction in the design of vehicles such as ships, planes, submarines, as well as propulsion systems such as scramjets and gas turbines, improving efficiency, extending capability, and providing more flexibility in mission planning.
“There have also been instances where sudden changes in the wake behind the decks of ships have compromised the operation of naval helicopters,” Bisetti also noted. “This is a very big deal to the Navy and the models generated by reinforcement learning could help them with simulating accurately complex flow dynamics at those conditions.”
Many industries and applications facing the challenges of unpredictable turbulent flow stand to benefit from these potential developments, including emerging fields like urban air mobility, which needs dependable models to help small drones safely navigate the gusts that blow through city landscapes.
Other team members working on the project include:
- Petros Koumoutsakos of Harvard University, an expert in AI and fluid dynamics, will oversee the theory and application of reinforcement learning;
- Charles Meneveau of Johns Hopkins University, an expert in large-eddy simulation and turbulence modeling;
- Greg Eyink of Johns Hopkins University, a mathematical physicist focused on the theory and modeling of stochastic systems;
- Mirko Gamba (Ph.D. aerospace engineering 2009, The University of Texas at Austin) of the University of Michigan, an expert in experimental methods for turbulent flows, will implement and operate the von Kármán turbulence simulator.
