Simulation in AI models revs up software-based auto designs

The increasing vehicle design intricacies and tight production schedules require automotive engineers to adopt new tools and techniques in order to build a differentiated automotive product. Enter artificial intelligence (AI) models for vehicle systems and R&D workflows.

Take the case of the state of charge (SOC) estimation for the battery management system (BMS) in electric vehicles (EVs), where algorithms receive sensor data coming from the battery to measure parameters like voltage, current, and temperature. The BMS uses this input to protect the battery while using specific techniques to perform SOC estimations.

Figure 1 AI models are now used to ingest the voltage and current data and make a prediction about the battery state of charge. Source: MathWorks

Automotive engineers now collect a bunch of data from these batteries—test lab data as well as roadside data—and train AI models that can make SOC estimations. Here, the job of AI models is to ingest the voltage and current data and make a prediction about SOC. “AI models are doing just as good of a job as previous techniques,” said Seth DeLand, data analytics product marketing manager at MathWorks.

Simulation data in AI models

Data preparation is the crucial first step in AI model-based designs for automotive applications. Data is paramount because AI models are trained based on the data. However, a model-based automotive design requires testing in multiple scenarios. On the other hand, automotive engineers gather data from operational fleets that they merely use as a starting point; otherwise, they don’t have data from different scenarios.

For instance, there could be an environment that is too harsh to collect data. “So, automotive folks are using simulation to generate those types of data in different scenarios,” DeLand said. “It’s much cheaper with simulation and much safer as well.” He clarified that simulation data doesn’t completely replace input data; instead, it’s being used to augment the roadside data engineers already have from the real world.

Figure 2 Simulation data is now being increasingly used to augment real-world data for training AI models. Source: MathWorks

Going back to the SOC estimation example, once automotive engineers have developed the AI model, they’d want to test how it performs with the rest of the BMS software. “That’s where simulation comes in,” said DeLand. “You want to simulate that control algorithm and see how this model performs in real-world conditions.”

He told EDN that model-based design has become vital in several automotive designs, as there is a lot of emphasis on ship left. Here, simulation enables engineers to model the software, controllers, and the rest of the environment to see how a design performs before moving to hardware.

“Simulation tools have evolved to support various stages of this development cycle as design engineers move closer to the actual hardware,” DeLand added. In other words, simulation enables engineers to move things earlier in the design cycle and have this toolchain of workflow that supports smooth transition when they move to the actual vehicle hardware.

Simulation challenges in AI models

The above information establishes the simulation’s influence and impact on the design of AI models within the automotive industry. But what are the best practices that simulation brings to the development of automotive AI models? What are the common challenges automotive engineers face when implementing simulation into AI modeling?

According to DeLand, AI models are often developed by other teams, not necessarily engineering teams. “So, you need to convert the model and port it from one toolchain to another,” he said. “The solution is automatically importing them from one AI framework to another.”

The other challenge is that after deploying an AI model, you might realize that you have new data and need to create a better AI model while carrying out a good update to the model, DeLand added.

Nevertheless, simulation is playing a vital role in the integration and testing fronts of automotive designs by providing engineers with a virtual environment. That allows automotive engineers to quickly try out different things—design ideas, tradeoffs, and more—in a low-cost way before going to the hardware stage.

DeLand adds that in terms of reducing hardware cost, one of the key applications we see emerging at the intersection of AI and simulation is using AI as a virtual sensor. The battery SOC estimation is just one example. “Rather than add another sensor to the vehicle, you can estimate that value from other sensors you already have in the vehicle.”

Figure 3 The notion of virtual sensing—sensing with software as opposed to adding hardware—is opening some interesting new opportunities in automotive designs. Source: MathWorks

MathWorks is currently working with a company that wants to estimate exhaust gas systems in the tailpipe. “The company wants to do that with an AI model while using sensor data available in the vehicle,” DeLand said. “That’s an opportunity to have one less sensor, which is a significant hardware save.”

Future of simulation in automotive AI

Still, automotive design is a diverse landscape, and the question is which automotive applications could benefit most from using simulation in AI models. According to DeLand, the places where simulation has been most helpful are places where there is a significant amount of software in the vehicle. That includes powertrain and automated driving.

“At MathWorks, we’d a lot of focus on powertrain controls on the automated driving side, and one of the things that are getting attention here is scenario generation, tools that build actual traffic scenes,” he told EDN. That’s because, in perception algorithms, there are so many things that you want to test. You need virtual ways of building these environments; otherwise, it will be too expensive to go out and collect this driving data.

DeLand noted a couple of important trends regarding the use of simulation in AI models serving automotive applications. “We see a lot of automotive engineers learning about the technology, taking online courses, and figuring out novel ways to bring AI to their workflows,” he said. “So, we will see AI proliferating into different automotive applications.”

Second, what engineers are simulating continues to grow complex as they try to increase the scope of what they are simulating. “It’s not just how a component interacts with other components or systems, and oftentimes it could be computationally very intensive,” DeLand noted.

So, the notion that engineers want to do more in simulation will not go away, he commented. Here, DeLand pointed to the role of AI-based approximation of different parts of automotive designs.

For years, engineers have used the phrase “reduced order model” to describe when they have some good physics-based model. They want to build something that’s an approximation of it because they want to expand the scope of what they are simulating. At the same time, they don’t want to wait for that simulation to be completed.

Figure 4 Automotive engineers are increasingly using AI to approximate designs. Source: MathWorks

“We are seeing a lot of interest in reduced order modeling as automotive folks are using AI to approximate some of the physics-based environment models they have done in the past,” DeLand said. “This is an area that’s likely to grow very fast.”

<!–
googletag.cmd.push(function() { googletag.display(‘div-gpt-ad-native’); });
–>

The post Simulation in AI models revs up software-based auto designs appeared first on EDN.

Simulation in AI models revs up software-based auto designs

Become a member

Become a subscriber

Become a sponsor