From aerospace to AI: what turbulence taught me about machine learning
I started in computational fluid dynamics at Sharif and ended up shipping production AI in Italy. The path is less strange than it sounds — both are about modelling systems too complex to solve exactly.
People are surprised when they see an aerospace-engineering master's degree at the top of an AI engineer's CV. They should not be. Computational fluid dynamics and machine learning are the same discipline wearing different clothes: both take a system too complex to solve exactly, and build a model that is useful anyway.
My M.Sc. at Sharif University was on turbulent premixed combustion. Turbulence is the canonical hard problem in physics — a system with so many interacting scales that direct simulation is impossible for any real geometry. So you do not simulate every eddy. You build a sub-grid model that captures the statistics of the small scales and feeds them back to the resolved simulation. I spent two years adding a Linear Eddy Model to CANTERA, an open-source combustion library, across C++, Fortran and Python.
That work taught me three habits I use every day as an AI engineer. First: respect the data-generating process. In CFD you cannot fit a model that violates conservation of mass; in ML you cannot fit a model that ignores how the data was actually produced. Both punish you for pretending the world is simpler than it is. Second: validation is the whole job. A combustion model that matches one experiment and diverges on the next is worthless; so is a model with 95% test accuracy that fails on the cases your business actually cares about. Third: the model is never the deliverable. In research the deliverable was a paper and a reproducible pipeline. In industry it is a shipped feature someone opens on a Tuesday. The model is the easy 20% in both.
The transition itself went through a Data Science master's at the University of Padova — the formal bridge from continuous physics to statistical learning. But the real bridge was the mindset. An engineer who has spent years modelling turbulence is not intimidated by a neural network. It is just another function approximator for a system too complex to write down. The difference is that the data replaces the physics — and you have to be twice as careful, because the data does not come with conservation laws to keep you honest.
If you are a physicist or an engineer wondering whether your background transfers to AI: it does, and more directly than you think. The specific equations do not carry over. The discipline of modelling a system you cannot fully see — that is the entire job, and you have been doing it all along.