My great expectations of AI after ChatGPT burst upon the scene had dampened in the years afterward. Design and engineering software companies were disappointingly slow to react, at first saying they had AI all along — for example, topology optimization or automatic dimensioning — then offering underwhelming and incremental enhancements.
I was reminded of a prisoner’s refrain, but applied to AI instead of sex, it becomes:
The AI I need I’m not getting, and the AI I’m getting, I don’t need.
But after talking to Gustavo Navarro, my mood brightened considerably. I had the pleasure of hosting Gustavo, founder of Divergence AI, on our Future of Design and Engineering Software podcast. Gustavo has left his native Chile to come to the Bay Area. He extends a hand to engineers, offering them an AI-assisted solution.
You might be thinking, “Isn’t everybody offering AI now?” But Gustavo is no opportunist jumping on the AI wave. He has a PhD in mathematics from UC Davis, and his knowledge of RF design is a mile deep. We need to pay attention to Gustavo, not because he has an AI copilot for Ansys HFSS, admittedly a niche solution, but because his copilot could serve as a model for all sorts of simulations.
CAE: It’s Complicated
Software companies have given engineers tools for design, simulation, and manufacturing — CAD, CAE, and CAM applications. They have not been smart tools. It’s still up to the engineer to supply the smarts, the creativity, to think of a shape, for example, then to use CAD, CAE or CAM to design, analyze or manufacture it.
CAD, CAE, and CAM tools all have their unique language and syntax. As with any language, they all take a while for the user to learn, and longer to become proficient. So complex are some simulation programs that few are able to master them, unable to take the classes, to get the degree, or train themselves in their use.
When it comes to simulation, a design engineer may be intimidated and avoid it, risking mistakes, fully aware that big mistakes can be frightfully easy to make. For example, type one extra zero for Young’s modulus and your factor of safety will be inflated by 10. Your bridge will collapse. Your plane will fall from the sky.
Simulation tools can be general, like stress analysis, specialized, like nonlinear deformation, or multiphysics with phase change. Honestly, how many of us can use them with confidence? Ansys HFSS, which stands for High Frequency Simulation Solver, seems to me, a mechanical engineer, the most complicated of all. HFSS is used to analyze the beam paths of antennas, among other radiation simulations. A radio frequency (RF) engineer would be a typical user, but not all RF engineers are trained in its use. Even with initial training, as with users of all sophisticated engineering programs, unless it is used frequently and regularly, it requires getting back up to speed.
That is the CAE predicament: the more specialized a simulation program is, the fewer of us who can use it.
Enter, Divergence AI
Gustavo has an answer to the CAE predicament: implement AI to provide a natural-language interface to HFSS. For this purpose, he has created Divergence AI. As an example, Divergence allows you to type “give me a beam plot for [an antenna] and plot it for every 30 degrees” instead of arcane HFSS commands. The AI will translate your prompt into PyAnsys code that HFSS will understand.
PyAnsys is the Ansys gateway to Python, the language of AI developers. It’s still a programming language, however.
So, thanks, Ansys, but I still don’t want to be a programmer.
The intersection of HFSS users and programmers who can code with PyAnsys is vanishingly small, says Gustavo.
Divergence AI, however, will allow the ordinary RF engineer to become an HFSS power user in relatively short order.
Using HFSS as the solver and Divergence as the interface underscores a vital point: engineering tools can be made easier with the right application of AI. What if we could run the most sophisticated of software just by “talking” to it in our language? Have it understand us instead of us understanding it?
Enhancing the Engineer, Not Replacing Them
Gustavo didn’t set out to replace HFSS. In fact, he reveres it. During our interview, he described HFSS the way a violinist might describe a Stradivarius — temperamental, demanding, and capable of extraordinary things in the hands of someone who knows what they’re doing. But, like a Stradivarius, not many people can just pick it up and play it.
“I’ve seen brilliant RF engineers get stuck before they even begin,” he told me. “It’s not because they don’t understand electromagnetics. It’s because they don’t speak HFSS.”
This, in a sentence, is Divergence AI’s thesis: engineers shouldn’t have to speak HFSS, or any simulation tool. They should speak in engineering, and let the machine translate.
And in Gustavo’s case, it really does translate — from casual English to PyAnsys commands, to structured Python, and finally to the mesh, solve, and post-processing routines that HFSS provides.
I asked if even an engineer such as me, a mechanical engineer, with little knowledge of RF and none of HFSS could use Divergence AI to do a simulation.
You still have to know what you’re doing, said Gustavo, though not in those words. He is far too polite.
“You still need engineering judgment,” he said. “AI is terrible at judgment. But it’s very good at busywork.”
It was indeed a sensible caution. Too many AI startups pitch a world where engineers merely describe a problem (“Give me a 3D model of a heat sink that cools this chip to 37°C and weighs less than 50 grams”) and the AI produces an answer. Gustavo flips this. He sees engineers as staying in the driver’s seat, with AI acting as the world’s most attentive grad student — the kind who shows up early, stays late, and never complains about re-running a model because you forgot to switch the units from inches to millimeters.
When Gustavo demonstrates Divergence AI, this becomes obvious. He types a sentence — literally one sentence — and HFSS performs the setup that would normally require clicking through a half-dozen dialog boxes, navigating cryptic menus, and knowing which parameters need to be defined before others. There’s something magical about it the first time you see it. But the magic is not that the AI is intelligent. The magic is that it has read everything and remembers everything, all of the commands necessary to execute HFSS given simply your intention.
Opening HFSS to the Many, Not the Few
RF engineers are not rare. HFSS power users are. If there is a trend in the ratio of the two, it is not heading in a good direction. As antennas proliferate — in phones, cars, drones, medical sensors, factory automation, everything — the demand for electromagnetic simulation is rising, not falling. Yet the barrier to entry holds firm.
“Companies end up with bottlenecks,” Gustavo said. “There may be only one person in a company who can run HFSS properly, so everything waits for that person.”
It’s an all-too-familiar story told in too many companies: one ABAQUS superuser, one shock-and-vibration guy, one person who understands acoustics, one experienced in turbulent flow with phase change. You can’t blame the companies. How many of these specialized experts can they find? How many can they afford? Applications like Divergence AI could eliminate that bottleneck. Not by dumbing down specialized tools or replacing them with AI guesswork — each would be a mistake — but by enabling every engineer to use them.
Gustavo told me about a customer who had been trying for weeks to set up a seemingly simple antenna sweep. Not because the physics was difficult, but because the HFSS workflow required a sequence of operations that the engineer didn’t know. They typed the request into Divergence AI, and the system not only generated the PyAnsys code but also annotated it, teaching the engineer what each block did.
“That’s the part I insisted on,” Gustavo said. “We’re not just doing the task — we’re showing the engineer why it works.”
This is where Divergence steps beyond being a copilot. It becomes a tutor, a mentor, a teacher who whispers best practices in your ear instead of shaming you for not knowing them. It is making engineering software behave as if it remembered the last fifty years of office wisdom — the stuff that used to be passed down in person, from senior to junior engineer, but which AI can now capture at scale.
From Chile to the Bay Area — and Why It Matters
Gustavo’s journey began in Chile, far from the R&D labs of the big defense contractors and big aviation companies, removed from the labs and anechoic chambers that are second homes to RF engineers in the US and wealthy NATO countries. His interest in mathematics — pure, abstract math, not the practical stuff of engineering, however, knew no boundaries.
“Math gives you humility,” he said during the interview. “You can be very smart and still not understand. That’s important for AI, because AI pretends to understand everything.”
In Chile, he didn’t have access to the software and hardware like HFSS until his doctoral work at UC Davis. It was then that the language of simulation software became his daily vocabulary. And even then, he found himself frustrated — not with the physics, but with the interface.
“It shouldn’t take 30 steps to do something that only requires three concepts,” he said.
Not feeling the frustration were fellow researchers and engineers. They sensed the friction but had gotten used to it. They assumed it was normal, a rite of passage. Gustavo didn’t. He saw an opportunity to build something that democratizes simulation, as CAD democratized drafting.
Why This Matters for the Entire Simulation Industry
When you talk to Gustavo, HFSS is the first chapter, not the last. Divergence AI is not a point solution; it’s a template. If you can translate English into PyAnsys, why not translate it into APDL for Ansys Mechanical, into the scripting language behind COMSOL, into Abaqus input files, into LS-DYNA keyword cards, into Star-CCM+ macros?
And once engineers become accustomed to commanding simulations the way they might direct an assistant — clearly, directly, and with full expectation of speed — the whole world of advanced simulation will light up.
There will be pockets of resistance, of course. Some simulation experts who have paid the price of learning how to use complex tools will not want their experience devalued. They will list all the usual fears, the planes falling from the sky, the bridges sure to collapse…
But engineering must march towards the accessibility of sophisticated tools. Like us all getting calculators, once a $400 accessory for the affluent, getting personal computers when the elite had access to the mainframes, the advancement of the profession as a whole is more important than job security for a few.
Engineers using an AI assistant for access to specialized simulation tools is a key point. Gustavo’s AI assistant is not hallucinating results. He is not building a simulation that “looks” right but may not be, with a mystery solver or no solver at all, an uneducated guess based on similarity. Divergence AI relies on HFSS doing what HFSS has already proven it can do: be a robust solution for a radiation solver.
The Road Ahead
I had to ask Gustavo if he thought at all about guard rails that would keep engineers from accepting results as gospel, even if they might be wrong.
He had.
“Verification is the key to trust,” he said. “If we speed up simulation setup, we must also make verification easier. Engineers must be able to trust what comes out of an engineering application.”
Until then, we consider Divergence AI a big step in the right direction. By extending the reach of HFSS to the engineers who need it but don’t have the time or patience to learn it and be proficient in its use, it is an excellent example of responsible AI assistance.
— — — — — —
Listen to the whole podcast with Gustavo on Buzzsprout here, but if you want to see the entire thing, including the demo, on YouTube, click here.