Fermilab scientist wins DOE award to solve AI's domain shift problem in physics
Aleksandra Ćiprijanović, an associate scientist at Fermi National Accelerator Laboratory, has received a 2025 DOE Early Career Award to develop artificial intelligence tools that address a persistent problem in high-energy physics: models trained on simulated data perform poorly when applied to real experimental data.
The domain shift problem occurs when machine learning models learn patterns from one dataset but encounter different data in the real world. In physics research, scientists typically train AI models on simulations because they're cheaper and faster than collecting experimental data. But simulations contain approximations and unknowns that don't match reality, causing trained models to fail when deployed on actual experimental results.
Ćiprijanović encountered this problem nearly a decade ago when she began applying AI to her research. She decided to build a solution that could work across the entire field.
A universal framework for physics data
Her project will create a software package designed for broad use across high-energy physics. Researchers will be able to upload their own datasets, select an AI model type, and specify what problem they want to solve-without needing to develop custom solutions for each application.
The framework will have modular components offering multiple options for data handling, model selection, and task types. Ćiprijanović plans to start with cosmological data, her primary research area, then test the framework on collider physics and neutrino physics data.
"I really do want to make a software framework that will be used across different high-energy physics frontiers," she said.
Why Fermilab is the right place
Fermilab's computing infrastructure and its concentration of experts across multiple physics disciplines make it uniquely suited for this work. The lab employs scientists from cosmology, collider physics, and neutrino research-the exact communities that will test and refine the framework.
"Luckily, at Fermilab, we have experts from all these frontiers," Ćiprijanović said. "Fermilab is the place to do this."
The DOE Office of Science Early Career Research Program distributes funding annually to support early-stage scientists at universities, national laboratories, and research facilities. The program has operated since 2010.
For researchers working with machine learning and experimental data, understanding domain shift is essential. Learn more about AI for Science & Research and how these techniques apply across different research domains.
Your membership also unlocks:
