Agentic AI in CAE: A Practical Experiment with Cosmon Argus
Imagine an AI agent that can interact with your CAE software, prepare geometry, generate mesh, define simulation parameters, execute workflows and help organize the results.
Sounds futuristic?
I recently had an opportunity to experiment with Agentic AI for CAE using Cosmon's Argus agent. I wanted to understand what this technology could actually do in a real engineering workflow, where it still needs human intervention, and whether it can genuinely make a CAE engineer's job easier.
What's In this Article
What is Agentic AI
Beverage can drop analysis
Results and observations
Will Agentic AI replace CAE Engineers
In this article, I'll share my experience, walk through a practical project and discuss what I believe Agentic AI could mean for the future of engineering simulation.
For this experiment, I worked on a fluid-filled Beverage Can drop-impact analysis. The objective was to study whether the can would survive an accidental drop and understand which factors influenced its structural performance.
Let's explore what I learned.
What is Agentic AI, and how is it different from Generative AI?
Most engineers are already familiar with Generative AI tools. We use them to understand technical concepts, write scripts, troubleshoot errors, prepare reports and find information. But there is a difference between an AI tool that tells you how to perform a task and an AI agent that can actually perform parts of that task.
Consider a simple example: You ask a conventional AI chatbot how to prepare a finite element model for a drop-impact simulation. It may explain the modelling steps, suggest suitable elements and provide recommendations.
An AI agent takes this a step further. Depending on its software access and capabilities, it can interact with engineering applications, prepare the model, execute commands, review errors and continue working towards the requested objective. This ability to plan and execute multiple connected tasks is what makes Agentic AI interesting.
It does not mean that the agent understands every engineering problem perfectly. Its performance depends on the instructions, available software integrations, data and engineering constraints.
For CAE engineers, the real question is not whether AI can answer simulation-related questions. It is whether it can take responsibility for parts of the actual simulation workflow.
The experiment: Can a beverage can survive a drop?
For this experiment, the objective was to understand how three factors influenced its ability to survive an accidental drop:
Drop orientation
Type of liquid inside the can
Can material
The problem was more interesting than it initially appeared. The can undergoes structural deformation during impact, while the liquid inside moves and interacts with its walls. This makes it a fluid-structure interaction problem which was carried out using Abaqus explicit.
The AI agent selected Smoothed Particle Hydrodynamics (SPH), a mesh-free method suitable for representing the large movement of the liquid during impact. It then prepared the simulation model, including the can, fluid and impact surface, while working within the available software's 1,000-node limit.
Four simulations, one interesting discovery
Instead of running just one simulation, I wanted to understand the influence of each factor independently. The agent prepared four configurations:
Case | Orientation | Fluid | Can material |
1 | Flat | Beer | HDPE plastic |
2 | 45° tilted | Beer | HDPE plastic |
3 | 45° tilted | Water | HDPE plastic |
4 | 45° tilted | Water | Aluminium |
The first simulation served as the baseline. Each subsequent simulation changed one major factor while keeping the other conditions consistent. The results revealed something interesting.
In the flat-drop configuration, the peak stress was approximately 2.9 MPa, below the assumed HDPE yield strength of 8.6 MPa. The can remained elastic.
When the orientation was changed to 45°, the peak stress increased to approximately 10.2 MPa. The plastic can exceed its yield strength and developed permanent deformation. That's around 3.5 times the stress observed in the flat-drop case!
Changing the fluid from beer to water made very little difference. The results were almost identical.
However, when the plastic can was replaced with aluminium, the can survived the same tilted drop. Its peak stress was approximately 43.9 MPa, well below the assumed aluminium yield strength of 275 MPa.
The biggest learning? Drop orientation had a much greater influence on the plastic can's survivability than the type of liquid inside it.
This was an interesting observation that became clear through the controlled comparison of simulation results.
What impressed me about working with the AI agent?
The simulation results were only one part of the experience. A few things about the agent's approach stood out.
1. It selected the modelling approach
The agent independently identified SPH as a suitable method for representing the moving liquid. This was important because selecting the right modelling approach is one of the first decisions in any simulation project.
2. It managed multiple iterations
Preparing four simulation configurations manually would involve repeating several modelling activities. The agent handled much of this work, changing the required parameters while keeping the other major conditions consistent.
3. It diagnosed a solver failure
During the initial runs, the solver encountered a segmentation fault. The agent traced the problem to an inappropriate tensile-failure keyword associated with the fluid definition, removed it and successfully executed the model.
This was probably the most interesting part of the experiment for me. Instead of simply reporting an error, the agent investigated its possible cause and attempted a correction.
Can we trust the results?
This is where I believe engineers need to remain careful. Although all four simulations completed successfully and the results showed physically consistent trends, the model had limitations. The available 1,000-node restriction meant that the shell mesh was relatively coarse, and the fluid was represented using only 132 particles.
The material properties were representative values rather than experimentally measured data from actual production material.
Only one drop speed and a rigid impact surface were considered. Therefore, the results are useful for understanding the relative influence of different factors, but the absolute stress and force values would need further validation before making a production decision.
A refined model, material testing and comparison with physical drop tests would be the logical next steps.
The agent also helped save an estimated 1 to 2 days of engineering effort. But I wouldn't consider speed the biggest benefit. The ability to prepare and compare four controlled simulation cases with less manual effort was more valuable.
Will Agentic AI replace CAE engineers?
I don't think this is the right question to ask. A CAE engineer does much more than prepare a model and execute a solver. We need to understand the physics, select appropriate modelling assumptions, interpret results and decide whether the simulation represents real-world behaviour.
AI agents may take over several repetitive activities, but that doesn't automatically remove the need for engineering knowledge. In fact, the ability to question an AI-generated model and validate its results could become even more important.
My takeaway
This experiment changed how I look at Agentic AI in engineering. The interesting part wasn't simply that an AI agent could create and run a simulation. It was that it could work through multiple engineering activities, handle an unexpected solver issue and help identify a result that wasn't obvious at the beginning.
At the same time, the limitations reminded me that automation and engineering judgement must go hand in hand.
Perhaps the future of CAE isn't about engineers doing less engineering. It's about spending less time on repetitive tasks and more time solving engineering problems.
What do you think? Can AI agents become a regular part of a CAE engineer's workflow?
I'd be interested to hear your thoughts.
Comments