Generative AI has made running an FEM analysis faster
With the rapid progress of generative AI such as ChatGPT and Claude, the scripts and interfaces needed to drive open-source analysis tools like FEniCS and OpenFOAM can now be assembled quickly by people who are not professional programmers. Only a few years ago, learning to operate the software was itself a barrier to entry. That barrier is steadily coming down.
This change deserves a straightforwardly positive reception. When anyone can get hold of an analysis environment, the base of the technology widens.
What has not changed: building the model, and judging the result
The value of FEM analysis, however, was never in being able to operate the software. In practice the genuinely difficult work lies in three stages.
- Building the model: which parts of the physics to account for and which to simplify, and how to set boundary conditions and material properties
- Judging whether the results are sound: determining whether the numbers that come out reflect real physical behaviour, or are artefacts of numerical error or a misconfiguration
- Translating the results into design decisions: turning the analysis into actual design changes and cost judgements
None of these is automatically taken over when AI performs the computation. However much you accelerate a calculation built on a mistaken model or the wrong boundary conditions, what comes out is only a wrong answer arrived at faster.
Not a profession losing work to AI, but one whose judgement matters more
At Mathematical Physics Labo we read the change this way.
The further AI drives down the cost of running an analysis, the more the specialist's role shifts from performing the computation toward judging whether that computation can be trusted and carrying it through to a decision. In other words, this is not a profession that loses its work to AI. It is a profession whose judgement becomes relatively more valuable as the cost of execution falls.
Through particle physics research at Fermilab (Fermi National Accelerator Laboratory) and SLAC (Stanford Linear Accelerator Center), the director of Mathematical Physics Labo spent many years training in exactly this: constructing physical models, and verifying computed results against experimental fact. Making active use of AI tools while bringing expertise to bear on the final judgement is where MPL's consulting sits in the age of AI.
How to start a conversation
"Could we run FEM analysis in-house now that we have AI?" "Is there still a point in commissioning it externally?" If you are weighing questions like these, do get in touch. One workable arrangement is for AI to be used well on the execution side while Mathematical Physics Labo takes on the parts that require expertise: building the model and judging whether the results hold up.