Who Cares if AI Violates CFD Physics? The Design is What Matters.
We all know that automation in case setup is already here. For standard cases, there’s almost no need for meshing or physical model setup—just a couple of clicks and you get results! However, know-how is still required to set up new physics, and I don’t expect AI/ML tools to reshape that much in the near future. In reality, though, we don’t spend significant time in product development setting up new physics, since that work is usually handled by research teams.
So the question is: how will AI change the CAE workflow in product development?
I believe the biggest opportunity lies in optimization.. In my view, the immediate opportunity for AI is in post-simulation optimization through CAE. I actually think this is more impactful than even a digital twin. In the near future, I believe every system and part that is manufactured will be optimized—at least based on the knowledge available at that time—thanks to AI. That is the true opportunity in product development.
The biggest question, however, is: what should the learning task be for the AI model?
In standard CAE optimization, you need to parameterize properties or geometry. You run an initial DoE, then start the optimization search. Advanced algorithms build surrogate models to accelerate this search. AI/ML research has focused heavily on improving surrogate models, and these methods were used in optimization workflows even before the AI boom. The shortcoming of all these approaches is the need for parameterization and the limitation of the search space to the chosen parameters.
I believe Generative AI can bring a lot of value here—and this is where we’ll see major impact soon. The value comes from a learning task that captures knowledge from previous simulations, not from directly replicating the exact physics represented by CAE models in an ML architecture.
As shown in the figure below (which I found online), I don’t want an ML model that reproduces exact CFD results. What I want is a model that captures the bounds and main correlations needed for my objective function. That’s enough for optimization. Even if the learning task violates some physics—so what?
Imagine a foundational model for external aerodynamics of a car. Through past simulations, it already knows the main enablers for reducing drag coefficient for a given style, and has learned correlations that influence those enablers. Importantly, it doesn’t need to learn mesh cell by mesh cell to achieve this. When you bring your vehicle CAD into this model for an optimization task (say, drag reduction), it will interpret the style, geometrical relationships (not the exact geometry), and the current status of your objective function. Then, the ML model could propose the top 5 candidate CADs that are potentially optimized.
Next, your CFD tool can test these candidate geometries, feeding results back to the ML model. The model then updates its knowledge of the gaps between suggested and CFD results. Through iteration, ML narrows the search until the constraints are met. This is why I think architectures that don’t rely on a one-to-one mesh model will win out. I’ve seen approaches where geometry is tokenized for the learning task, and I have high hopes that this is the way forward.
In my opinion, forcing ML models to perfectly respect physics or replicate CAE models is a waste of time—we’d just be reinventing the wheel. However, applying physical guardrails may help ensure the extracted knowledge stays in a meaningful range.
In a nutshell: I don’t see much value in AI/ML tools that aim to replace CFD tools. Instead, ML can bridge the gap between analysis and optimization, especially in complex applications where hundreds or even thousands of parameters influence the design.
As I mentioned in a previous blog post, I don’t expect a general AI/ML model that can solve all CFD problems. Useful generative AI models will be application-specific. Still, we may converge toward shared techniques for tokenization—similar to what we already do in meshing.
