Early this morning, SimScale presented a webinar to debut what it calls “the first fully integrated and cloud-native AI-based physics predictions in simulation software.” The German-based engineering software company offers a completely cloud-native simulation platform via a software-as-a-service (SaaS) licensing model. All an engineer needs to access it is a browser and an internet connection.
The AI features are available to all SimScale users at no additional cost. SimScale partnered with Navasto to add AI to its online simulation platform. Navasto is also based in Germany, and offers AI technology that tackles engineering applications. Here are some take-aways, first impressions and thoughts on this announcement and the future of AI and simulation.
Takeaways of SimScale’s AI-Simulation Tool
SimScale sees AI as the future of simulation because they see it as a way to meet current engineering challenges, including:
- Optimizations to reduce costs, energy/material use and carbon emissions.
- Generative design.
- Product complexity.
- Mass customization.
- Digital twins.
The idea is to use AI to get quick results so that engineers can perform more simulations to meet these challenges. Instead of running simulations on thousands of generated designs, engineers run simulations on a hundred designs, use that data to train an AI machine learning (ML) model and then assess the remaining generated options. Once the model is trained, AI can act as an alternative to solvers such as FEA, CFD and other simulation techniques.
SimScale can now replace its simulation solvers with AI technology in any workflow it offers. As the platform is completely cloud native, so are the AI tools. And just as engineers can run a SimScale simulation on a browser, share it and collaborate with team members in real-time, the same is true for these AI models. Alternatively, APIs are offered to connect AI functionalities into third-party workflows.
The AI models are managed by SimScale users and customers. Though they are pretrained to a certain extent, reports SimScale to engineering.com, they still require some training from the user. SimScale offers workflows to train ML models using project and legacy data. However, a bring-your-own-data approach is used to train the models. The upside is that the models will be optimized to the users’ specific needs.
Some simulation examples demonstrated with the AI technology include:
- Vehicle impacts.
- Vehicle aerodynamics and drag optimizations.
- Centrifugal pumps.
- Stress and strain.
- Heat transfer of a CPU heat sink.
First Impressions of AI in Simulation
According to Steve Laine, application engineering manager at SimScale, “AI will do for engineering what the Internet did for engineering,” and revolutionize it. I do not doubt this to be the case. However, I see SimScale’s current iteration of AI-based simulation as a first step along that exponential curve. This gives the company a head start against competitors, but it comes with a significant drawback in the form of needing to train all the AI models. In essence, engineers require some ‘activation energy’ to get the AI-reaction started.
"AI methods have proven to be able to use past simulation data to drastically accelerate future engineering workflows,” says David Heiny, CEO and Co-Founder at SimScale, to engineering.com. “Yet the complexity of data collection and preparation as well as model training and deployment oftentimes hinders adoption. We believe a cloud-native integration of such AI methods right next to proven PDE solvers and deploying both of them with the same UI and API will be a big enabler for broader AI adoption in engineering organizations."
To put it another way, when you input a geometry and pre-process it for the first time, the AI tool can not give you an estimation of the part’s structural integrity. To do that you need to run FEA models first. It is only after you produce dozens, maybe more, simulations that an AI model can be produced. Then it can be used to replace the simulation solver in the workflow.
In this sense, the AI tool is used in a similar fashion to a response surface modeling (RSM), or similar 1D/2D simulation simplifications. Though Jakob Lohse, Product Manager at Navasto, suggests that the AI tools will be more accurate than these simplifications, he admits their ability to offer fast results based on minor changes to a model and geometry can replicate the new workflow from Navasto and SimScale.
The presenters were also cagey as to how much data would be needed to train a model, but it was later reported to engineering.com that good results can be made using about 30 simulations. Though the bring-your-own-data method does offer users the flexibility to customize ML models to their personal use cases, it means they need to have, find or make that data in the first place. That isn’t always an easy ask when racing to market.
Richard Szoke-Schuller, product manager at SimScale, explained that simple examples, like some used in the presentation, needed around 200 simulations to train the model. As more boundary conditions, inputs, complexity and physics are added to the mix, the amount of training data only goes up from there. If, however, some of the AI-models were fully pretrained, under the hood, on common simulations, this might not be the case.
The Future of AI and Simulation
In a final slide, SimScale reviewed its vision of the future of AI and simulation over the next year or so. This is no doubt a skeleton development map of its future offerings.
It appears that many of these offerings anticipated my first impressions, the most impressive being an AI-solver equivalent to a PDE solver. With this, engineers could theoretically swap out their traditional solvers on a first assessment of a new design—assuming the PDE solver was fully trained and certified for the given use case. This would be huge, as it would enable engineers to still tailor their models to their needs during pre-processing, but they would not have to run hundreds of simulations to eventually get the AI benefits.
SimScale also talked about the possibility of self-training and self-managed machine learning models. These could learn as they go based on a user or organization’s habits, assessments, workflows and product lines. In theory, these could automatically offer up all the simulations and assessments an engineer would need as soon as a geometry is drawn up.
Finally, SimScale dreamed up how its AI tools could be used in a digital thread. No doubt with the self-training models, these could learn from all the data an organization has—even what’s outside of SimScale. For instance, a model could learn from live production data that a part isn’t being manufactured properly. In theory it could then automatically offer up suggested solutions as soon as the issue is flagged.
We can mark 2023 as the year that AI and simulation tools really started to mingle in a meaningful way. However, it’s clear that there is much more on the horizon. Perhaps in 2024, I can ask J.A.R.V.I.S to run a simulation on my jet-boots and it won’t just do all the work on its own but will also give me three suggestions to improve what I already designed. SimScale might have a leg up on this, but they too have a way to go.