2025-08-26: Summer in the Midwest

 

Wrapping up my internship with a presentation at Fermilab

I was fortunate to spend the summer interning at Fermi National Accelerator Laboratory under the supervision of Dr. Marc Paterno. The whole experience was full of firsts. I had never been to Chicago or even Illinois before, and I had never seen a bison or a particle accelerator in real life. These are the kinds of things you only hear about on the news back in Europe. 

At first, the dorm I was placed in felt a bit cramped. But over time, I got to know many interesting people from all sorts of backgrounds, and somehow my room didn't feel small anymore.

How PSO finds the global minimum

For my project, we developed an efficient GPU-based multidimensional global optimization algorithm, which we submitted to HiPC 2025. It's a novel hybrid method that has two main phases: particle swarm optimization for global search and then a local refinement phase with BFGS, and automatic differentiation. It outperforms existing methods in terms of both convergence and runtime. It was equal parts a challenge and an interesting discovery process.

A practitioner using a global optimizer does not know the landscape of the function, which may contain many local minima. The animation shows how PSO behaves for a highly multimodal 2-dimensional Rastrigin function. The color of the contour shows the height of the function. At first, the particles are spread randomly across the search space. Then each particle's movement is influenced by its personal best-known position but is also guided toward the global best-known position, which then moves the entire swarm to the global best solution. The arrows refer to the velocity vector of each particle. We can observe that, by iteration 20, most of the swarm found the global minimum located at (0, 0) in a search space where there are 121 local minima, so each yellow circle is one where a gradient-based method can get stuck in.

During my internship, I learned a tremendous amount about experimental setup and software engineering practices, including working with larger codebases, writing test suites, and how to explore and visualize data with R (the animation above was made using ggplot). These are skills I’ll carry beyond this summer. 

Beyond the technical work, I had the chance to visit several of Fermilab's experiments, including the famous Muon g-2, Mu2e, and NOvA experiments, and the Short-Baseline Neutrino Detector.  When I say "visit", I mean actually going down into the mineshafts where the particle detectors live. The MINERvA experiment is currently using the detector I'm pointing to in the photo. Standing there made me realize how much mystery still surrounds us and how exciting it is to even play a small part in exploring it.

~ Dominik Soós