Typing this in 2023, it would have been hard to guess five years ago that AI has become the craze that it has become. ChatGPT and Bard can amazingly replicate human writing patterns and converse with you. Stable Diffusion and Midjourney generate very realistic images just from a prompt (like the viral selfies from history). All are doing new things in unique ways and certainly mark a change in how some things will be done from here on out.
And then we come to the case of CFD solutions from AI. Putting these two terms in a sentence together is fine, but the range of combinations of CFD and AI goes from exceptionally and uniquely useful all the way to redundant and absurd.
What Is AI? An Unsolicited Non-Specialist View
Before describing the role of this tool, I want to give some rigor to what it is. The way I think of AI is that it is basically the most flexible input-output system there can be. Any number and type of inputs can be converted to any number or type of outputs. A properly trained network is the most powerful and flexible curve fit you can think of, on steroids. An 800 pixel x 800 pixel array of RGB values that you want to turn into a word, like cat, dog, or horse, telling you what animal is in the picture? Neural networks have you covered.
For applications to fluids problems neural networks can do better than other input-output relationships mainly because they can handle well the non-linearities of the relationships we encounter. The classic example that comes to mind is the drag crisis shown below.
Are you going to fit a polynominal to this (from Wikipedia)?
As much as the whole thing appears like (and mostly is) a black box, there are comprehensible lessons one can take from certain canonical examples, like the undercomplete autoencoder. Some of it is comprehensible, but there are known problems with hallucinating outputs that aren’t real, and in general when the cards are down people tend to follow a trust but verify approach to what a neural network-based model is telling them.
For Surrogate Modeling
Neural networks have been used for some time to supplement CFD results – namely, they can serve as a way to create a surrogate model of a design space. As shown in the drag crisis figure above, this is really a place where these models can exceed the capabilities of other approaches. A code primarily for turbomachinery, Numeca, has a neural network surrogate model as a key component of the optimization routine within it, and I have used this model in the past to very good effect for torque converter optimization. Some initial CFD runs were carried out (I was told the rule of thumb was at least 6 CFD runs per factor, DOE form) to train the neural network. An optimization routine was then applied to the surrogate model to guess the optimal performer based on the target criteria I told it. After the optimization routine identified what it thought were the parameters of this best performer, the CFD case with those inputs was run to check. This process could be repeated until no more performance improvement was found in successive runs. For this application, neural networks worked very well.
For Generating CFD Solutions
Here’s where things get weird, and you have to start to think about what you believe. On the one hand, the Navier-Stokes equations are the statements of physical truth whose solution we seek, and performing a process by which they are solved seems like the most simple – or really the only sensible – approach to take. On the other hand, approaches to solving steady flows already make use of the fact that, hey, as long as you end up with a solution which conserves mass, momentum, and energy, who cares how you got it? Some AI black box that churns out something which is then found to obey the laws of physics as we like would be just fine, practically.
As far as this full-feature CFD replacement AI use case goes, we find lots of marketing talk about solving RANS problems in much less time than the CFD process would take – conveniently omitting mention of the training time involved. One of the reports I saw required something like 100,000 CFD simulations in the training set to arrive at a “time-saving” AI model. What can’t a good CFD analyst do with 100,000 simulations that an AI model can?
Ultimately, I think a large part of my reaction to this is driven by some (small) degree of fear that this whole AI thing is going to wash over the existing CFD processes like a tsunami and leave all of the traditional science and numerical folks useless. Some folks I work with now come in, almost weekly, with a list of new terms that I haven’t heard before and don’t understand, and my overall impression is that neither do they. There’s a real risk AI is just the topic-du-jour to justify many mid-manager jobs and we’re all getting dragged along. If we can hang on through this period of jargon and their lack of understanding of good engineering, and lack of respect for physics, we may be able to come out on the other side with some good offshoots having come from this phase. Part of me is fearful that 98% of all of us will be left behind, as noted above. But my experience also tells me that the fundamentals never stop being important.