Aaron Wienkers

I have spent a decade taking the same problem apart in different disguises. It began at Berkeley in astrophysical fluid dynamics, chasing turbulence in eccentric accretion discs, and the object of study has changed many times since while the method has not. Take a complex nonlinear flow, resist the urge to settle for a correlation, and find the physical mechanism underneath it. That instinct, formed on protoplanetary discs and sharpened at Livermore, is the thread through everything that followed.

At Stanford I read for an MSc in flow physics, writing high-performance Fortran with sixth-order compact finite differences, and took the two courses (CS229 and CS230) that grounded the machine learning I now use daily. The Cambridge PhD moved the same machinery into the ocean: how symmetric instability at submesoscale fronts mixes down the thermal wind and drives a vertical exchange that global climate models are still too coarse to resolve. A short spell at Bern followed, building an ocean biogeochemistry sub-module into the GFDL Earth-system model, coupling it to the existing atmosphere and land components at HPC scale.

The pivot happened at ETH. I lead the EERIE ocean–atmosphere evaluation across four major European climate centres, and somewhere in the middle of it the question turned over: from running the physics models to interrogating the AI ones now moving in to replace them. The scepticism I built characterising real turbulence is exactly what tells me when an emulator's output should not be trusted, and that is the through-line into industry. I build and validate the weather and climate AI that other people are going to bet on.