AI in Engineering Simulation: The Most Pressing Questions Answered

Introduction

We've all heard about AI transforming industries, but what does that really mean for engineering simulation? Is AI just another trendy term, or is it truly transforming the way engineers predict, optimize, and innovate? As AI continues to enhance predictive modeling, improve accuracy, and drastically reduce simulation times, many questions arise about its real-world applications and limitations.

To answer these pressing questions, James Shaw, Founder and Managing Director of Fastway Engineering, shares his expert insights on how AI is impacting engineering simulation in this exclusive Q&A session.

You can also watch this discussion in video format here.

Now, let’s dive into it!

What role does AI play in engineering simulation today?

That's a great question. I mean, it's a loaded question. We've had AI in engineering, and if I take a step back for a second, what do we mean by AI in engineering? For me, that just means: do we have a predictive model? Right? AI is just the latest sort of algorithm sets, data sets, and approach for predicting the future for any design. And so we've had the ability to take a bunch of information, test data, simulation data, put it together, interpolate some solution, and say, hey, this is an optimized solution. 

So today, what I see in AI is, hey, I've got this historical data, I've got generational sort of prior designs and how well they've done reliability, performance, etc. and I want to come up with the latest design, new materials, new manufacturing processes, whatever. How is it going to perform? I mean, that's like the engineer's dream. That's what we're trying to do here, and so the latest AI tools facilitate that. The million dollar question is how. How do they do that? That's a little bit longer answer.

Why is AI becoming so critical in simulation workflows and how is Ansys integrating it into its tools? Can you give us an overview of the key AI-powered solutions in the Ansys portfolio?

So a couple of things to unpack. The first question was how is AI being used? And today, the way that we see it and the way that we and our customers are using it, first and foremost is in GPTs, right? 

So we have customers up until last year, I would say if you had a question about FEA, or CFD, and you asked students or young engineers, “What would you do if you need to learn something in the middle of your workday?” the answer would have been to Google it. Up until 2023, or maybe even through last year, people would have said, "I’ll go to Google, type in my question, and see what happens.” What happens now is it sends you out to the internet, and you're getting a bunch of noise. I don't wanna say junk, because there's some good information out there, but you gotta sift through the junk to get the good stuff.

Well, if you ask an engineer now, you ask a student in college right now “Hey, if you're trying to learn ANSYS, you're trying to learn FEA, CFD, or simulation techniques, and you have a question, what do you do?” And there's going to be a split now between Google and chatGPT. Future engineers, they don't care. They're going to just go to whatever search bar is closest to their mouse button, right?

We're starting to see these search bars get embedded in the tools. So right now, we've got Ansys GPT. Ansys GPT is a large language model. It's at your fingertips. You just open it up and you type in: Hey, I'm trying to do this thing in Ansys Mechanical, and it will automatically search all of the databases that we have. So that's the first implementation of AI, right? All these large language models, I'm typing in text and I'm getting answers.

It's like having your own personal assistant, except for Ansys GPT, you have a personal assistant that is completely trained in mechanical, fluent, and the entire portfolio. They're instantly tapped into our learning material, our tech support material, the help files, everything instantly. That's a huge value, because if you're stuck and you're learning, the last thing you want to do is wait and the very last thing you want to do is find a bad answer and have you go down a route that's wrong, which we see with the junk that's out there on the internet right now. 

But that's just the starting point. The starting point is how do I aggregate all this technical data and get it easy to find at my fingertips? Really, what we want to do is we want to embed AI into our simulation math. Not just a tech support tool, but in the simulation math. 

What we have in our portfolio right now is we have an AI implementation that's centered around big data. What do I mean by that? Instead of doing one simulation, I do 10 or 100 or 1,000 and so, we have a lot of statistical math that we wrap around our physics. There are ways to optimize within that data set and so we've got some AI algorithms that allow us to do that. That is what I would call sort of interpolating AI, right? 

Let's talk about extrapolating or generative AI. Generative AI is saying, “I’ve just fed you a data set. You now understand the relationships between my inputs, outputs, and the physics involved.” From there, I can give you new geometry or present a "what-if" scenario that’s outside the data you’ve already seen - something that’s extrapolated from the training data. That's our SIEM AI tool. There are a few different ways you can use AI in simulation, but I think the jewel in the crown, if you will, would be the twin AI tool, which is basically saying, “I've got a predictive generative AI tool that I'm running on the fly and not only is that fed with high quality training data, say historical, but I'm actually feeding it live on the fly with sensor data from a machine or from the field.”

It doesn't get any more accurate than that.

Ansys TwinAI: Real-Time Simulation and Predictive Engineering Model

Ansys SimAI claims to deliver results in seconds instead of hours - how does it achieve such drastic speed improvements?

Everybody wants to know! I mean, first off, the relationship between speed and accuracy has always been, a little, a little bit of a battle, right? That engineers don't trust an answer if it comes too fast, right? Because we kind of know the real world is nonlinear. The real world is full of noise and it's sometimes really hard, right? All models are wrong, some are useful, and that's because models are simplified versus the real world.

If you take a model and you simplify it so much that it becomes seconds to solve, engineers don't want to trust it. So, how do we do that? Well, of course, as we mentioned, training data. You want to feed it dozens or even hundreds of prior sets of data. That can be a simulation, a trustworthy, full physics simulation. That can be some test data, right?

But what do we do with that?  I'll give you an example of CFD. So I plug in my geometry. I have a couple million cells. A couple million cells inside of that takes hours to converge. So, I've now weighted and plugged in all this information, but I've got a reliable answer, and I know it's going to match test data. It takes me whatever, hours, right? Reality is once you've converged that, and you've got a full field of temperatures, pressures, velocities, densities, and the reality is, if I did a sensitivity study on the inputs and understood how much they impacted the outputs, I could actually take this full field, millions of cells, and break it down to a very, very simple equation. I could reduce millions of degrees of freedom down into a pretty simple relationship. My in is this and my out is that, and that's what we call a reduced order model.

So, I'm reducing the order of the problem. The order started off as millions of degrees of freedom. Now I'm reducing it to five, right? If that reduced order model is accurate, I can now predict the answer of millions of cells and thousands of iterations of CFD in a matter of seconds because, I have a trustworthy relationship, a mathematical relationship between my inputs and my outputs. That's how it's working. That's how it's so fast. 

One major challenge with AI models is getting reliable training data. How does Ansys tackle this issue?

That's the million-dollar question. A great source of training data-there are several-is full physics simulations.

We've got customers who say, “Hey, I don't have historical data. I got a design out in the field, the engineers have all retired, the new engineers have no idea how it works, the test data is written in pen and paper and locked in a cabinet somewhere, or it accidentally got thrown out. ​​How do I build up training data on historical designs?”

I know that they're in the field. I know they're reliable, but that's not enough. A reliable design is not enough to feed into a SIEM model. Reverse engineer that, put it in a full physics model, FEA, structural dynamics, CFD, heat transfer, whatever, and solve it. Get that model to match your test data and do it again, and again.

This involves generating new data points: new temperatures, new flows, new forces, new displacements, new temperature profiles. By doing this, you develop millions of data points with spatial discretization, like 3D CAD files, and temporal data, such as transients. When this data aligns with the existing test data, you can fill up hard drives with gigabytes of data for training. 

On the other hand, test data alone is not full field. When I make a heat exchanger, I don’t know the full temperature field. I don’t have enough thermocouples, and I certainly can’t afford that many. I might have half a dozen thermocouples, or maybe an IR camera, but even then, I’m dealing with their own limitations and errors. Still, I feed that data in. So, you really need a mix of training data. But when it comes to the quantity of training data, that's going to come from simulation, no doubt.

How do you see AI evolving in engineering simulation over the next decade?

The next 10 years in simulation are really all about AI. What I mean by that is the development of predictive models that are smarter and may require less training data. Additionally, as a society and engineering community, we’ll start building public training data sets.

Now, companies don't want to share their training data. It was very expensive to build, and they don’t want to give it away. But I think at the industry level, if you think about some of the industry organizations like ASME, ASTM, and SAE - representing fields like mechanical engineering, aerospace, and possibly civil or architectural engineering - have a vested interest in regulating and ensuring the quality of engineering.

I believe that over the next decade, there will be initiatives from these organizations, or others, to make training data available - either publicly or behind a paywall for those willing to invest. Ultimately, there will be more accessible training data, which you can either purchase or download for free. This will significantly improve the quality of simulation, generative AI tools, and engineering as a whole.

The other side of the coin, of course, on the speed side is, hardware and the democratization of computing power. Of course, primarily, we're talking about CUDA cores on Nvidia GPUs and the whole technology. 

Watch our latest webinar with NVIDIA for more on Insights on GPU-based High Performance Computing (HPC)

While Moore’s Law may be dead, we have other technologies enabling incredibly fast computing. With large amounts of training data, both the quality and quantity are increasing. The speed of generative AI will continue to rise, making these tools more accessible. As generative AI and simulation AI tools become more accessible, they’ll be used more frequently. This means optimizations will improve, factoring in more variables, and designs are just going to be better.

Are there any misconceptions about AI and simulation that you'd like to debunk?

Great question. The misconceptions around generative AI for engineering and simulation are the same misconceptions that we've had in simulation for the last generation or two. That is, the belief of what the cost or the difficulty is. Again, these tools are changing about every six months, right? AI capabilities are changing at least -or say at most- every six months, possibly even faster than that. If you're not paying attention, then you're automatically going to have an outdated view of what the capability of these tools are, of what the cost of them are, and of what the complexity and the difficulty is. 

As we just continue to move forward,the cost is going to go down and I'm talking total cost, not licensing costs. Software companies have invested interest to increase their licensing costs, presumably over time. At the end of the day, the total cost of using these tools - factoring in human labor, hardware, and training data - will continue to decrease. At the same time, the capabilities will keep improving. I think some misconceptions out there stem from misalignments, like reading a blog a few months ago or hearing something in a podcast (self-promotion aside) that you believe, but never verified. Ultimately, to stay ahead, you need to stay at the cutting edge.

Talk to people, talk to variety of people, get your new sources, get your technology sources from multiple areas. Don't just hone in on one particular source and do it often. Otherwise, you're going to have an outdated opinion.

What advice would you give engineers hesitant to adopt AI in their workflow?

Advice for people who are hesitant to start to introduce AI: 

First off, your hesitation is usually driven by a lack of understanding. ّّIf you don't fully understand the tool, you should be hesitant. This was similar to the early days of simulation - people questioned whether the stress predictions were accurate. Well, yes, they were, because the math had been proven and the tools were reliable. The same is true for AI now.

Do your homework, get your research from multiple sources, not just a person who's got a vested interest in selling you something or somebody who has old data.

I think the other hesitation is people are fearful to introduce something because they don't feel like they have the bandwidth to learn or that they're going to introduce a technical risk into their next project. For those people, I would offer the following advice: 

Whenever you start a new project, push back on your management and company by saying, “This project should include some bandwidth to explore new tools.” Just like you account for unexpected risks - like adding an extra week or hour to a project - you should also factor in the effort to try new tools. Engineering is all about managing risks, and that includes the risk of not adopting new technologies. Make room for this potential in future projects.

A lot of times we see companies say, many companies try to isolate a special project in a vacuum to introduce a new tool, aiming to compartmentalize the risk to their business. But that’s not the most realistic way to introduce AI. Instead, you should incorporate AI into a program that can tolerate a small amount of technical risk. If the AI doesn’t deliver as expected, you can always fall back on more traditional methods that you know work, whether that’s full physics simulations, hand calculations, trial and error, or physical prototyping.

For those who are hesitant, it all starts with when you're initially scoping out projects moving forward. Whether they’re CapEx or OpEx, talk to your project manager and leadership, and suggest factoring in an extra couple of percent for experimenting with new tools. I think that's going to reduce the hesitation and the risk, which is the most important part for these businesses.

Conclusion

ُIn short, AI's role in engineering simulation is undeniably transformative, with its potential to change how engineers    approach design, optimization, and problem-solving. 

As we continue to see AI evolve, particularly through platforms like Ansys and generative models, the barriers to adoption will diminish, with cost and complexity continuing to decrease. For engineers hesitant to embrace AI, it's important to recognize that the risk of not adopting new tools may outweigh the challenges of integrating them - because in engineering, sometimes the best solutions come from the most unexpected designs.

________

Have other questions about AI in engineering simulation or how to integrate it into your workflow? 

Contact us today