AI not only improves designs and predicts anomalies, but it can also streamline workflows to get there.
Quite a lot of AI buzz in the design and simulation world focuses on optimizing and predicting outcomes. For instance, there are endless stories out there about how AI can optimize a final design or predict a maintenance cycle—and those are important topics I’ll touch on here. However, not a lot of the AI buzz focuses on how it can optimize the workflows we have to achieve those outcomes.
Fatma Kocer, VP of Engineering Data Science at Altair, is vividly aware of how AI can be used to improve the workday of a CAE user. “When we setup to use the newer machine learning (ML) techniques, what we focus on is how can we leverage data science, which includes machine learning, to improve our processes and outcomes,” she said.
Altair splits its AI strategies into three categories:
- AI augmented simulation (or ML-augmented CAE): The use of AI tools to improve workflow processes.
- AI powered design: The use of AI tools to optimize products and explore the design space.
- Predictive data analytics: The use of AI to assess and test the performance of products and designs in the field. These assessments can also be fed back into the design process.
AI Augmented Simulation is best implemented under-the-hood of the software tools engineers use. The best examples will feel natural and will be so embedded into the workflow that you may not be able to notice them—unless you know where to look. As a result, the driving force behind its implementation is software developers such as Altair.
So, why are developers interested in using AI to optimize their software workflows? The answer is that by simplifying an engineer’s job, it grants them more time to utilize and implement other tools—such as AI powered design and predictive data analytics.
“I’ve spent years increasing the use of these techniques to come up with better products and I’ve always complained that people were not jumping in on it,” Kocer explained. “But as we look in the data science, and where to leverage it, it’s clear why people are not jumping on these technologies. It’s because people are spending a lot of mental effort in comparing the model. They don’t have the remaining energy [or time] to look into what happens when you have the model, or how do you get a better design.”
The Data Challenges Engineers Face Implementing AI Powered Design
Kocer is correct; it isn’t easy to implement AI systems into a design process. It requires a lot of energy, time and a significant budget. For an example, let’s investigate the challenges posed by something as simple as legacy and current data systems and formats.
Kocer notes that this challenge can be summed up in the term “data discipline” and that there are a few ways organizations can have issues with their data discipline. First off, simulation data itself can be a challenge because it isn’t in a format compatible with AI system workflows.
“We’re going to be working with 3D shapes, CAD files and finite element models,” Kocer said. “Those are not in a format that is digestible by machine learning algorithms. Machine learning algorithms like structured, clean data. Usually, our data is clean but not structured. So, one part of our investigation that we’ve worked on is, how do you convert 3D geometry into structured data that can then be used by machine learning algorithms.”
Another issue is the storage and organization of data. For this issue, Kocer notes there might be two camps.
“A lot of companies have data management systems that collect and organize data. I know, when I visit a major OEM, they will have a system they have been using for a long time,” Kocer said. “But it may not have all the functionality we would require to extract, clean and filter the data for machine learning. They’re not built for those requirements in mind. But regardless, the data is somewhere and mostly organized.”
“Beyond OEMs,” she added, “in smaller to medium sized companies, they usually dedicate some desktop, cloud or multiple clusters for data to be collected. They would have to be organized and cleaned.”
Kocer explained that the reason a lot of design organizations don’t have data discipline is because they tend to have only one thing in mind: producing good designs. Once a design is settled upon, the previous designs, along with the path to get there, are often forgotten. In fact, because simulation data is so large, unless the data is needed for some certification or regulatory reason, it might be deleted.
“If you’re not using a data management system, you’re not in the habit of extracting the metadata that could be useful for machine learning,” said Kocer. “When I think about data discipline, I don’t mean keeping the data somewhere. I mean saving the data and keeping it organized, just like a closet, so when you need it, you can find what you need.”
Clearly, challenges like this make it difficult for engineers to implement AI-powered design. That’s why Kocer and her team are focusing on using AI to optimize Altair software workflows; to give engineers the time to tackle these issues.
Examples of AI Augmented Simulation in HyperWorks
So, what would AI augmented simulation look like? Let’s look at some of the tasks associated with model pre-processing—notoriously one of the most tedious and time-consuming parts of the workflow.
One of the first thing engineers do when they input a model with thousands of parts is to start grouping those parts into clusters that will receive similar meshing treatments. Kocer uses the analogy of opening a grocery bag. The first step is to group the groceries into what goes into the fridge, freezer or cupboard.
“This clustering takes a long time because you have to literally go one-by-one, or maybe come up with a script,” she said. “We now have part clustering in Hyperworks. It is an unsupervised machine learning functionality. With the clustering algorithm, we group each part in the assembly per their shape. So, in each cluster you see the same exact, or similar, parts. Then you can go to each cluster and mesh them, as opposed to meshing or idealizing each part.”
AI augmented simulation doesn’t just optimize an individual’s workflow, it also can optimize the workflows of entire organizations. Kocer noted, “Traditionally, simulation has been a trial-and-error process where the learning from past designs is all done by the engineer, and it all stays in the engineer’s mind. It becomes company information for as long as the engineering is there. Whereas AI augmented simulation is a way to capture that learning process, in terms of machine learning methodologies, that can be transferred from one engineer to another, one department to another. This way the data lives in the company and improves with the company.”
For instance, if a company continuously creates parts or products that look a certain way, an AI augmented system would notice this. Then, perhaps it could start suggesting geometry, at the very start of the design process, based on a machine learning algorithm taught by previous designs.
Let’s put it this way: People tend to recognize an Apple product from a mile away, regardless of what that product is. Wouldn’t it be useful for an Apple design engineer, who is designing the next new whatever, to input a few parameters and have an AI system come up with a handful of ‘brand appropriate’ geometries? This is much faster than starting from a blank slate.
It’s no wonder that with AI augmented simulation, engineers will have more time on their hands. Will they use this time to implement other AI tools, as Kocer suggests? I hope so. But what would you do with that time? Comment below.
