How Well Does Your FEA Program Know You?

The infusion of artificial intelligence (AI) and machine learning (ML) is swiftly reshaping engineers’ simulation workflows. Traditional simulation software users rely on the robust physics solvers to predict and analyze the deformation, displacements, pressure distribution, heat buildup, airflows and fluid flows in their design concepts. This time-tested approach has a heavy computing cost, which translates to time and money.

The new approach made possible by AI is to use legacy simulation data to train and develop an AI-like algorithm that can predict simulation results from design shapes, similar to how a veteran finite element analysis (FEA) or computational fluid dynamics (CFD) user might develop intuitions based on experience. This novel approach, along with support for natural language in user interfaces (UI), could spawn a whole new way of assessing design concepts.

Your Past Simulation Data Is a Goldmine

In January of this year, Ansys officially launched Ansys SimAI, described as “a user-friendly artificial intelligence platform that combines the predictive accuracy of Ansys simulations with the power of generative AI.” Unlike a typical FEA or CFD program, it doesn’t come with any solver. Instead, it offers an environment that lets you develop an AI-like predictive algorithm.

“SimAI is an AI platform that can take any physics simulation solver as a black box solver: FEA such as Ansys Mechanical, or CFD like Ansys Fluent, or Electromagnetics like Ansys HFSS,” Prith Banerjee, chief technology officer of Ansys, clarifies. “You can give it a CAD model and some boundary conditions, and train the AI model. Once it is trained, it can simulate other CAD models and other boundary conditions much faster—in seconds versus hours.”

SimAI, according to Ansys, “provides data-driven insights rapidly and reliably by learning from your legacy simulation data. The user experience prioritizes simplicity; no coding or data science expertise is required. Just upload your data, select relevant design outputs and generate your AI model.”

By feeding a program like Ansys SimAI your previous simulation results and the relevant CAD models, you are in essence teaching the program to figure out the correlations between the design topology and the FEA or CFD results. With thousands of processing cores, computers take only a fraction of the time it would take a human learner to process the data. In time, the AI extracts an algorithm that captures correlations. This algorithm is the faster, quicker model you can now deploy to solve future simulation tasks, bypassing the need for time-consuming physics-based simulation runs.

“Technically, SimAI can take experimental data from physical tests, as well as simulation data from rival simulation tools like Dassault and Siemens and Altair, and it will still work,” Banerjee adds. “However, there is currently no plan to integrate with partner or competitor software.”

Beware of the Scope of Your Data

For small and mid-sized engineering teams, the challenge with this approach is the limited amount of data. GM, Ford, Toyota and Airbus may have hundreds or thousands of simulation runs for the AI to digest and develop a reliable predictive model, but for many engineering firms, the archived data may be limited to dozens of simulation runs, due to the costly nature of simulation jobs. This means the usefulness of the trained algorithm is applicable only to a narrow scope of scenarios that closely resemble the sample data pool.

In a recent DE 24/7 webcast, Astrid Walle, research and development engineer, Siemens Energy, cautions, “If you fed the program data about airflows and pressure distribution on cars with narrow chassis, then use the algorithm to make predictions on cars with wide chassis, you won’t see accurate results because the algorithm is clueless about the new type of car it’s dealing with.” (“Is Your CFD Hallucinating? Unpacking the Opportunities and Challenges of AI for Simulation,” keynote to the DE 24/7 Design & Simulation Summit, October 31, 2024).

When available data is limited, you may augment the dataset with synthetic data, adding mathematical variations to the data with reasonable assumptions. With this approach, you may further bolster the accuracy of the predictive algorithm, or the reduced order model, by accounting for the inherent uncertainty. This is a specialty of software such as SmartUQ, described as “a modern AI and Uncertainty Quantification tool optimized for engineering applications including simulation, digital twins, testing and manufacturing.”

I’m Your Virtual Assistant

Seth Hindman, Autodesk’s director of Product Management and Strategy, envisions use of natural language in the company’s products. He says, “Why do we have to use terms like inertia relief? Why can’t it just be unconstrained? What do engineers really want to ask? They want to ask: Is the part going to fail? Where will it fatigue? Can it lift twice as much weight? Can more liquid flow through it?”

The good news is that ChatGPT-style chatbots are becoming mainstream. At Autodesk University 2024 in San Diego, CA, Autodesk CEO Andrew Anagnost says, “As we move to simple [UIs], what could be simpler than natural language? That is what we’re doing with Autodesk Assistant.”

The new Autodesk Assistant lets you ask questions using natural language, and is currently trained to provide answers based on the Autodesk knowledge base, covering the company’s product portfolio.

Jeremy Stadtmueller, Autodesk’s director of Product Management, Design and Engineering Documentation, says, “Engineering tools, whether it’s Autodesk Fusion or those from our competitors, are inherently complex. AI’s advantage is, it gives us a way to remove that complexity.”

Up to now, in many design and simulation tools, the user needs to know the correct terminology to search for a tool, or the location on the menu to access the tool (for example, the Spline tool), which results in weeks or months of training.

“In the future, I can envision that if you say, I’d like to cut a pocket, Fusion might recommend a set of tools, or ask you a set of questions, like, what RPM (revolutions per minute) should it be, what materials [does it use], then execute the command for you,” says Stadtmueller.

Tools like Autodesk Assistant will not only be trained on industry best practices and Autodesk’s product knowledge, but also on the personal usage of individual users.

“Our first release of Autodesk Assistant is based on our own knowledge base and help documents. But we will expand to aggregate public knowledge from curriculums and research papers … We’re looking at customizing the data model for the individual users. It would then be aware of how you tend to build geometry, and your past projects that are similar to what you’re attempting to create, so it might even autocomplete some jobs for you,” says Stadtmueller.

In April 2024, Ansys released Ansys GPT, an AI-powered virtual assistant. This is the company’s test case for offloading first-tier tech support to a chatbot trained on the Ansys portfolio and knowledge base.

“AnsysGPT was developed both to enhance our customer support experience and free up time on our technical support team so that they can focus on other challenging problems and provide more advanced support,” Anthony Dawson, vice president of Customer Excellence at Ansys, says. “As beta testing progresses, we’ll get a better idea of the extent to which the service can reduce customer support workloads.”

Currently, virtual assistants like Autodesk Assistant and AnsysGPT are confined to scouring the database for answers and presenting the available solutions in natural language. But they’re also the harbingers of smart design and simulation programs that can deduce our engineering intent and propose suitable solutions.

More Ansys Coverage

More Autodesk Coverage