How simulation helps engineers overcome common challenges
As the demands on industry become more complex, engineers are starting to turn to AI to help them optimise their models.
By leveraging simulation to improve the data fed into a model, engineers can test for accuracy before utilising it in the real world. The process not only helps keep up to date with the latest demands, but it can also help to reduce costs and improve efficiency.
Electronic Specifier caught up with Seth DeLand, Product Marketing Manager at MathWorks for MATLAB machine learning and data science products, to talk about the design of effective AI models and the role of simulation in helping engineers overcome common challenges, and the role of AI in algorithm development.
Why is AI important? And why has it grown so quickly?
AI models enable computers to do what they do best, which is compute and process data and develop a model that can represent that data. At MathWorks, we focus specifically on the engineering disciplines and things like that.
AI models come out of the research community, and they're used to demonstrate certain applications. Then the engineering community gets a hold of them, and they come up with all these different new problems and applications that they can use them on.
It’s interesting to see is how these AI models have proliferated across [the] different areas of engineering and people have found new and novel ways to apply them.
There are a couple of things that have really driven [the use of AI]. [Predominantly] the large amount of data we've collected in the last several years ... Whether that is sensor and telemetry data coming from a vehicle, or data about customers, or data from their software simulations – AI models need data.
What other key demands are you seeing them being used for?
We see them used both on the design side and the production side.
On the design side, engineers have all sorts of workflows for designing products and services, and they use a lot of software simulation. They want to build models of the system that they're creating, and they want to simulate as much as possible before they build the real thing so they can gain confidence that their design is going to work before they spend the money that it requires to build a physical prototype.
On the production side, we're seeing engineers come up with ways to integrate AI models into their production systems. Sometimes that can be a topic that we call virtual sensing, where you have some quantity that you want to measure, but you don't necessarily have a sensor for it. An example of that would be an exhaust system on a vehicle. There's all these different chemical compounds and things you would want to keep track of, but you don't want to have a whole bunch of sensors [because] sensors are expensive.
Where we've seen the most adoption, so far, would be in automotive. There’s a lot going on in terms of more automated driving trends and electrification as vehicles become fully electric.
The aerospace industry is solving very similar problems. Medical devices are another one, you get biometrics and sensor data coming from medical sensors, [like] EKG and ECG. We’ve seen AI models [being used] to do analysis of the data and classification, which is exciting and another great example of where you can measure the data.
It used to require an expert to go from raw data to the final classification that you want. AI models are starting to get good enough to be able to do that as well [as the expert].
What kind of data can engineers get out of a simulation?
It depends on what they want. Theoretically, you can get anything you want.
You can go as detailed or as high level as you want when you're defining the simulation. It's up to the engineer to decide what trade-offs are important. Do they need something that's very, very detailed, but is going to take a long time to run, or do they want something that's more efficient, but is not going to be as accurate and is going to make a lot of assumptions. It's a tough decision to make because they're very strong trade-offs.
What are some of the challenges that engineers face when implementing AI modelling?
One [challenge] is a lack of AI expertise. [Engineers] are experts in their domains. They work on a specific system, whatever it is an aeroplane, a car or a medical device, and they are absolute experts in that area.
But they're not familiar with the AI techniques themselves. And it seems like there's new types of models every week. Trying to keep up with that when that's not your primary job is really challenging.
[MathWorks] is trying to make it easier for engineers by giving them standard model types that work well to a wide variety of applications [like point and click tools, giving them advice on how to manage data for training the model] rather than one specific research project that does well in one very narrow application.
[Also] there's so much data available, but at the same time for a lot of these engineered systems, there's not publicly available data. Companies don't want to share all the internals of how exactly their computerised systems are working.
What are the benefits of using AI before it is deployed in the real world?
The main benefit is catching issues earlier in the design process. When you're developing the thing that you're simulating, it’s something like 10 times less expensive than capturing the issue after you've built a prototype and are doing physical testing.
It's also much faster to fix, so it's a cost thing. It's [also] a time thing. You look at the timelines a lot of these companies are on … they're incredibly short compared to what they used to be. So they've been able to drive down that development time through model based design and simulation.
How do you see AI evolving?
I would say more types of models. The research committee is going to keep trying all sorts of different things. There's no limit to what can be done or what can be included in one of these AI models. At the same time, I think there's going to be more emphasis on making it easier to use so that domain experts, who aren't necessarily AI experts, can use 90% of the models.
So, lots more models, but at the same time making things easier to use and lowering the bar. Those are probably the two higher trends I would say.