Yakup Emre Şahin

I’m a doctoral researcher at the University of Liège in the Science with AI Lab (SAIL) led by Prof. Gilles Louppe, working on scientific machine learning.

I did my Master’s in Physics and AI at LMU Munich with a DAAD scholarship and wrote my thesis at Helmholtz AI in the group of Prof. Niki Kilbertus on symbolic regression for dynamical systems.

Before that, I completed a double major in Electrical & Electronics Engineering and Physics at Boğaziçi University, Istanbul, and did research internships at places including CERN and Koç University’s gravity group.

Email  /  CV  /  Scholar  /  Github

Yakup Emre Şahin

Research

Can machine learning learn the scientific models we don't have yet, rather than just emulating the ones we do or inferring their parameters?

That's the question behind my work. I build ML pipelines that learn scientific models directly from observations, combining generative models and Bayesian frameworks with latent world models, and probing those world models to see what physics they have actually learned. Right now I'm applying them to unfolding detector effects in particle physics. Weather and heliophysics are natural next targets.

If you find this question as mind-boggling as I do, feel free to reach out for a chat or collaboration!

Selected Publications


MIO overview: trajectories from several systems are encoded, aggregated per system, and decoded into symbolic ODEs
Predicting symbolic ODEs from multiple trajectories
Yakup Emre Şahin, Niki Kilbertus, Sören Becker
NeurIPS 2025 ML4PS Workshop

We introduce MIO, a transformer-based model for inferring symbolic ordinary differential equations (ODEs) from multiple observed trajectories of a dynamical system.

[arXiv] [code]
Marginal posterior probability density over the scalar-Gauss-Bonnet parameters
Subtleties in constraining gravity theories with mass-radius data
Ekrem S. Demirboğa, Yakup Emre Şahin, Fethi M. Ramazanoğlu
Physical Review D, 2023

Mass–radius measurements of neutron stars can constrain certain alternative gravity theories like DEF and scalar–Gauss–Bonnet models, though their effectiveness depends on the specific model and mechanism.

[arXiv]
Hollow electron beam imaged on the YAG screen
DC and pulsed electron beam test facility at CERN
A. Rossi, …, Y. E. Şahin, …, M. Wendt
JACoW IPAC 2023

A CERN test stand was built for the Hi-Lumi and ARIES projects to evaluate electron lens and space charge compensation components, featuring versatile diagnostics and instrumentation, with first results confirming the HEL gun design.


Psst: click anywhere on the page to place a star, then add a few more and watch them orbit.

Design stolen from Jon Barron's website. Source code

Gradients are due to a personal disease.