Scientific computing · CWI, Amsterdam
Small scales
Big questions
I’m Syver. I study turbulent flows, the details our simulations cannot see, and how to model what’s missing.
Explore the ideasIllustrated atmosphere
Your time integrator
is a filter
A simulation smooths the world in space. It does the same in time. What happens when the model accounts for both?
Read the interactive essayAnalytic signal · illustrative, not experimental data
02 Notes from the work
All writing ↗︎What symmetry buys a learned turbulence model
What do symmetry constraints buy a neural turbulence closure? Same accuracy floor, 25× fewer parameters, and one missing input for Reynolds-number generalization. An interactive companion to the paper.
Interactive essayYour time integrator is a filter
Forward Euler is not a first-order approximation. It is an exact statement, for a flux averaged over the time step. What a code leaves out is one more term in the closure problem, and that term is sixty years old.
Interactive essay · Part 3 of 3Nine components, not six
Every closure model you have used returns a symmetric, local stress. The exact one is neither. Both rules break for the same reason, and the reason is the shape of a grid cell.
Draft · Part 2 of 303 A little context
Between mathematics
and the moving world
I’m a postdoctoral researcher at CWI in Amsterdam. My work connects numerical methods, machine learning, and the physics of turbulence. Away from it: history, languages, and the piano.
More about meIncompressibleNavierStokes.jl
Differentiable, GPU-accelerated fluid simulation. The software behind the questions, and a place to start asking your own.
Explore the software