Hi, I’m Kozak.
I work on fusion physics, scientific machine learning, and the awkward gap between a good equation and code that runs fast enough.
At the University of Tokyo, I’m researching physics-informed neural operators for two-dimensional lower-hybrid-wave simulations on the TST-2 spherical tokamak.
Before Tokyo, I studied computational science in London and space science in Taiwan. I also build production software and help engineering teams decide when a learned surrogate is genuinely useful.
Making wave simulations fast enough to explore.
My PhD work asks whether a neural operator can preserve the physics of a lower-hybrid-wave solver while cutting the time needed for each run. The current scheme reaches 281.8× faster inference in a rectangular domain and 25.8× in TST-2 circular geometry.
Read the projectA few problems I’ve spent real time on.
Research papers, production experiments, and the numerical details in between.
Making lower-hybrid-wave simulation fast enough to iterate on
A physics-informed neural-operator surrogate for TST-2, with a 498 MB footprint and up to 281.8× faster inference.
PhD research · fusion plasma
Replacing hours of mold-flow simulation with seconds of inference
A Fourier neural operator for non-Newtonian fluid dynamics, tested against a production simulation workflow at under 3% error.
Industry R&D · operator learning
Teaching a neural network where the magnetopause ought to be
Reg-PINNs combine an empirical physics model with a neural network and reduce RMSE by roughly 30% over Shue et al. [1998].
Published research · space physics
Taiwan → London → Tokyo
The subjects changed along the way. The habit of moving between physics and software did not.
- Japan University of Tokyo PhD · nuclear fusion research
- UK Imperial College London MSc · applied computational science
- Taiwan National Central University BSc · space science & engineering
Have a simulation that takes too long?
Send me the solver, geometry, runtime, and the bottleneck you care about. I’ll give you an honest first read on whether scientific ML or classical HPC is the better answer.