Like many of my technology colleagues with interests in high performance computing (HPC) and digital engineering, I have been both excited and intrigued by the emergence of machine learning / AI in the research world of computational fluid dynamics (CFD).
It seems two relatively distinct communities have emerged – both highly experienced in their own disciplines, with quite different understandings of the opportunity for combining AI and CFD.
The fluid dynamics specialists
The first community – fluid dynamics specialists – has tended to regard with some scepticism the messaging from the AI community (informed most recently by IT-driven interest in replacing traditional methods with GPU-accelerated neural networks) and the examples that are offered.
This is not to say that the fluids community is not open to new ideas – far from it – but those ideas generally need to be rooted in solid physics.
Machine learning and AI evangelists
The second community – machine learning and AI evangelists – understands how to leverage the immense power of modern hardware, and the neural network algorithms that allow for the representation and interrogation of very high-dimensional surrogate models. This applies data science to reduce the seemingly huge complexity to a point where trends can be identified and meaningful decisions made.
Where do the communities intersect?
An interesting example of where the two communities appear to interpret the results of the same experiment differently is when an AI / ML model is trained on a set of aerofoil sections (and their corresponding CFD analyses) to learn the flow patterns and then reproduce them for a new aerofoil.
I have seen this exercise repeated several times, by different research groups around the world.
The fluid dynamics community does not give weight to success here. The aerofoil section family most commonly used is in fact based on only 4 or 6 independent engineering parameters. Therefore, a sensibly constructed design of experiments should allow it to be comprehended by only a few tens or hundreds of evaluations.
The AI community regards the ability to learn the flow patterns using only a few tens or hundreds of thousands of CFD solutions as evidence that their technology can successfully be applied to the most famously complex problems in physics – simply by scaling the method.
How can we bridge the gap?
Bringing these communities together is important for both to achieve success – which is one of the reasons that I am so optimistic about the newly created ERCOFTAC Special Interest Group (SIG) in Machine Learning for Fluid Dynamics.
Working out what does and doesn’t work when we apply AI / ML technology to fluid dynamics is vitally important. These workshops will help us identify the fluid dynamics challenges where new technology adds value and provide examples where the solutions are effectively implemented.
The European Research Community on Flow, Turbulence and Combustion (ERCOFTAC) is one of the world’s most respected and knowledgeable scientific organisations. It is well known for its high standards of quality and rigour via its online Knowledge Base or Best Practice Guidelines.
ERCOFTAC organised its inaugural “Machine Learning for Fluid Dynamics” workshop in March 2024 at the Sorbonne in Paris. The event, which featured a presentation by Zenotech’s Constantinos Vagianos, brought together around 160 participants from around the world to cover a broad range of topics where AI / Machine Learning and Fluid Dynamics intersect.
The workshop generated a real buzz, proving it arrived at just the right time, and I look forward to sharing new developments from this ground-breaking community.