Machine Learning Models Decode Proton Structure

Researchers at Argonne National Laboratory are using artificial intelligence (AI) to create detailed maps of proton structure, potentially accelerating discoveries in fundamental physics. Their machine-learning (ML) models analyze physics data to reveal how quarks and gluons are distributed inside protons and to search for physics beyond current theories.

The research addresses a long-standing challenge: understanding what’s inside protons, the particles found inside atomic nuclei. While scientists know protons contain quarks and gluons, mapping their precise arrangement has remained difficult.

The Argonne team’s AI and ML models analyze data from the world’s most powerful particle physics experiments, as well as data extracted from supercomputers using the techniques of lattice quantum chromodynamics. Their approach serves two distinct purposes and introduces innovations in how AI handles scientific uncertainty. The research team builds neural networks and related models that perform two functions. First, they map proton structure by analyzing physics data. One approach, called a variational autoencoder, reconstructs the distribution of gluons inside protons using theoretical calculations. The results agree with existing methods, confirming that such ML models can extract accurate information from data. Second, these models search for signs of undiscovered new physics that the prevailing Standard Model cannot explain.

The team also develops AI-based methods to identify when Standard Model theory predictions break down. The team created uncertainty-aware models that identify when data fall outside their range of validity. This capability helps researchers distinguish and explore unknown particle physics scenarios. The XAI4PDF method developed by the team shows which data features influence theoretical predictions, making analyses more transparent. Argonne theorists are now co-developing agentic systems like ArgoLOOM, an AI framework that can plan and execute physics calculations automatically. This and related approaches coordinate different computational tools to accelerate research in fundamental physics through new cross-frontier calculations.