I spent the last decade in the quantum industry, focused on silicon spin qubits and superconducting qubits. Far more of that work is about electromagnetic design than people expect, which is how I ended up spending my days in solvers. What pulled me toward computational EM and machine learning was the cost of the development loop. A single design iteration can mean hours of solver time. That quietly caps how many ideas you can afford to try. Heaviside, Arena Physica's forward-inference model, collapses that to milliseconds, and the value isn't just about speed. When iterations are cheap, you stop settling for the first design that clears spec and start looking for the best one. That's why I joined Arena Physica as a Research Scientist focused on the simulation side of our EM foundation model.
A model trained on physics is only as good as the physics you feed it, so much of my work sits upstream of training: understanding where our FEM solvers agree and disagree, why some classes of geometry are harder to simulate accurately than others, and what that means for the data the model learns from. Some of that is solver internals and numerical conditioning. Some of it is mesh generation, or designing experiments that isolate one physical effect from the three others it's tangled up with. The through-line is making sure that when a prediction is wrong, we can tell whether the model is at fault or the simulation that taught it.
I’m most excited about answering the generalization question. Simple, well-behaved structures were always going to work; the interesting part is whether it would work on dense, irregular, real-world hardware. I'm also excited about something less glamorous: eval engineering. Knowing exactly what our numbers mean, and whether the simulations that produced them are themselves right, is what makes every other claim of progress trustworthy. I like that Arena Physica treats this as real research rather than housekeeping.
There are two core reasons why I joined the team. First, Arena Physica is making the harder bet. A lot of AI-for-engineering companies use a language model to drive a conventional solver: the AI sets up the problem, but the solver still computes the physics, which quietly makes the solver the ceiling. Instead, Arena Physica is betting on a foundation model trained directly on the underlying physics itself, which is a much more interesting problem. Second, the problem rewards my background. A decade of quantum experience is fairly illiquid domain capital in most places, but here it converts directly.
My favorite memory at Arena Physica so far is nothing dramatic, but really the small stuff. Thirty seconds of cheeky banter at the water cooler. Someone landing a terrible joke while we're all standing around waiting for our salad bowls at lunch. Those are the moments I'd actually miss.
They're also why our technical work is successful. I've had meetings where a teammate and I sharply disagreed and we pushed hard on each other's reasoning. What struck me was that the idea won and not seniority or whoever spoke first. Walking out genuinely aligned, rather than one of us just conceding, is one of the best feelings in this job. It's a lot easier to get to that point with people you actually know and are in the office together with, plus the occasional evening out as a team. And last but not least, I'm looking forward to the company retreat this Fall.
I'm fairly new to the team, so weigh my tips accordingly if you’re interested in joining us, but I felt the signal was clear from my interview loop onward. Everyone I spoke with combined real technical depth with genuine curiosity about ideas that weren't theirs. Bring a real opinion, and expect it to be stress-tested! That's not a hazard of the culture, but the point of it. Nobody here is offended when you question a benchmark or an assumption. Instead, they'll help you design the experiment that settles it.
And in classic Arena Physica tradition of sharing a hot take, here's mine: quantum hardware progress has been gated more by microwave engineering than people think. As these systems have scaled aggressively over the last few years, a classical bottleneck has appeared. Each qubit you add brings its own control and readout channels into the same package, and the coupling paths between them grow quadratically. Full-wave simulation of that quickly gets expensive until you can no longer afford to simulate the thing you're trying to build. None of that is a physics problem. It's an EM design problem, and the cost of predicting it.
We're actively hiring for additional Research Scientists - you can learn more at arenaphysica.com/careers.