The New RF Engineer

Ask a room of RF engineers what the job requires and you will get a familiar list: electromagnetics, transmission lines, matching, noise, nonlinearity, a working feel for what a Smith chart is telling you, and years of intuition about why the measurement disagrees with the simulation. The list is still right. It has stopped being enough.

Artificial intelligence is restructuring RF engineering from the bottom of the stack to the top. At the circuit level, learned surrogates predict the response of a matching network or an antenna element in microseconds where a full-wave solver needs minutes. At the system level, reinforcement learning tunes beamforming weights and power-amplifier linearization while the link is running. At the policy level, machine learning is becoming the way regulators decide who gets which slice of spectrum, and when. Each of those layers used to belong to someone with an RF degree and a lot of patience. What the tools take is the middle. The middle of the job is the part where a designer knows what they want, knows roughly how to get there, and spends three weeks sweeping parameters in a simulator to close the gap. It is repetitive, well specified and expensive in human hours, which makes it the perfect target. Physics-informed networks, inverse design and automated optimization already do it faster and more thoroughly than any of us can by hand, and they are getting better every quarter.

I do not know anyone who went into RF because they loved waiting for a mesh to converge.

What the tools cannot take is the beginning and the end. At the beginning, someone has to decide which problem is worth solving, what a good answer would look like, and which constraints are real and which are habits. At the end, someone has to look at a result and know whether it is physically plausible, whether it will survive fabrication tolerances, whether the package will ruin it, and whether the measurement is telling the truth.

That judgement is what AI is worst at, and it only comes from having built things that failed. A model trained on a million simulations has never had a connector arrive from the shop half a mil out of spec. It does not know that the perfect 50-ohm match it found needs a via aspect ratio no board house will touch. An experienced engineer sees those problems in a glance. The glance is the new centre of the profession.

The optimizer can drive. Somebody still has to know where the car is going.

There is a subtler reason the ends of the job stay human, and it has to do with how these tools fail. A learned surrogate is fitted to the space of designs it was trained on. Ask it about something inside that space and it is astonishingly good. Ask it about something just outside, a new material, an unusual aspect ratio, a frequency band with no examples, and it will answer with the same confidence and no way of signalling that it is guessing. An inverse-design optimizer is worse in a specific way: given a model with a flaw, it will find the flaw and exploit it, presenting a geometry that performs beautifully in a simulator that is wrong. Neither tool knows what it does not know. The engineer’s job is to be the part of the system that does.

So the list grows rather than shrinks. The RF engineer of the next decade needs everything on the old list, plus some things that used to be optional.

Coding stops being a nice-to-have. Not because engineers will write the models themselves, but because the workflow now runs through scripts, APIs and data pipelines, and an engineer who cannot script is an engineer waiting for someone who can. Fluency with the tools follows the same logic: how to set up an inverse-design problem, how to constrain it, and how to tell when the optimizer has found a flaw in the model rather than a solution to the problem. Optimizers are very good at the first and will present it as the second.

Packaging and integration move from a specialist’s concern to everyone’s. At millimeter-wave the transition, the interconnect and the package are the design. An engineer who understands only the chip, or only the antenna, will be automated around by a tool that co-optimizes both.

Communication matters more, not less. When the technical middle compresses, what sets an engineer apart is the ability to frame a problem for people who do not share their background, and to defend a decision with an argument rather than a plot.

Education will have to move too, and it has been slow to. Most RF curricula still spend their laboratory hours teaching students to do by hand what the tools now do in seconds, and very few hours on the two things the tools cannot do: deciding what to build, and finding out whether what was built works. The reform is not to teach less physics. It is to teach it against real hardware. A student who has soldered a connector badly, watched the match collapse, and traced the reason on a network analyser has learned something no sweep can teach, and has acquired precisely the judgement that becomes the scarce asset in the next decade.

Hiring will follow the same logic. A transcript full of simulation coursework will say less and less about a candidate, because the tools that produced those results are available to everyone. What will say more is a record of things built, measured and shipped, and the ability to explain, in plain language, why a design that looked perfect on a screen failed in the world and what was done about it. The engineers with those stories will be the ones who stay expensive.

It helps to picture the working day. The new RF engineer opens the morning with a set of candidate designs an optimizer produced overnight, not a blank schematic. The first hour is spent throwing most of them out, for reasons the optimizer could not know: this one needs a via the board house cannot drill, that one has a resonance sitting exactly where the package will land, a third is a beautiful answer to the wrong question because the requirement changed on Friday. The survivors go to a solver for confirmation, and while the solver runs the engineer is on the phone with the mechanical team about a housing tolerance, and then on the bench with last week’s prototype, whose measurement disagrees with everything. By evening one design is worth building. That day contains almost no sweeping. It contains a great deal of judgement, and it is a better day than the old one.

For the people who run engineering organisations, the implication is uncomfortable but clear. Headcount plans built on the middle of the job, on the number of designers it takes to iterate a set of matching networks, are measuring the wrong thing. The scarce resource is the people who can tell a real result from a plausible one, and they are made slowly, by letting them build and break things early. An organisation that automates its junior work and keeps its juniors away from hardware is quietly cancelling its own senior engineers a decade in advance.

The risk here is not replacement. Wireless hardware is getting harder, not easier, and demand for people who understand it is rising. The risk is that engineers who located their worth in the middle of the job, the sweep and the tune and the iterate, will feel it evaporate underneath them while the beginning and the end go to someone else.

The beginning and the end were always the interesting parts. The tools are handing them back. Take them.