Incorporating deep neural networks into lepton-nucleus interaction modeling

To understand how the universe works at its most fundamental level, scientists often fire beams of particles at atomic nuclei and watch what happens when they collide. One of the most important questions in this field involves neutrinos, ghostly particles that rarely interact with matter but carry crucial information about the history of the cosmos. To learn from these elusive particles, researchers must reconstruct the energy of the neutrino beam by analyzing the debris left behind after a collision. However, this reconstruction is incredibly difficult because the target nuclei are complex, crowded environments where the rules of simple physics often break down. The uncertainty in modeling these nuclear interactions is currently the biggest obstacle preventing scientists from making precise measurements of the universe's fundamental properties.

To solve this, physicists often turn to electrons. Electrons behave very similarly to neutrinos when they strike a nucleus, but unlike neutrinos, their energy is known exactly before the collision. This makes electron scattering an ideal laboratory for testing how well our theories describe nuclear behavior. For decades, scientists have collected vast amounts of data on how electrons bounce off carbon atoms, creating a detailed map of what happens during these collisions. The challenge has been that while we have a perfect map of the total outcome, we lack a clear guide on how that total outcome is built from the specific, individual pieces of the puzzle. This gap has left a significant blind spot in our ability to simulate these events accurately for the next generation of experiments.

A team of researchers at the University of Wrocław has developed a new method to bridge this gap, effectively combining the precision of experimental data with the detailed storytelling of computer simulations. They created a system that takes a highly accurate prediction of the total collision outcome, derived from a deep neural network trained on decades of electron-carbon data, and breaks it down into the specific types of interactions that occur. This allows them to generate complete, realistic simulations of every particle flying out of the collision, ensuring that the total picture matches the real world while the individual details remain physically consistent.

The core of this work involves a clever two-step process. First, the researchers used a sophisticated computer program to simulate millions of electron collisions with carbon atoms. This program relies on established theories to predict the specific details of each event, such as whether the electron knocked out a single proton, created a pion, or triggered a more complex reaction involving multiple particles. These simulations provide the "story" of the event, describing exactly what the final particles look like and how they move. However, the total number of events predicted by these theories often does not perfectly match the real experimental data across all energy ranges.

To fix this mismatch, the team introduced a second layer: a deep neural network. This is a type of artificial intelligence trained specifically on the vast archive of real electron-carbon scattering measurements. The network learned to predict the total number of collisions that should happen at any given energy with extreme precision, effectively acting as a perfect reference for the total outcome. However, this neural network is only valid within a strict range of energy transfers: it cannot describe collisions where the energy transfer is too low (where the nucleus recoils as a whole) or too high (where the interaction becomes too complex). The researchers then took the computer-generated events and adjusted their importance, or "weight," so that the total number of events matched the neural network's perfect prediction within this valid range. This process ensures that the simulation is anchored in reality.

Crucially, the team did not just adjust the total count; they also figured out how to distribute that total count among the different types of interactions. They calculated the fraction of events that should be quasi-elastic scattering, where a single particle is knocked out, versus meson-exchange currents, where the interaction involves a more complex exchange of forces between particles, and single-pion production. By applying these fractions to the neural network's total prediction, they created a new, hybrid framework they call eWroNet. This system produces a complete set of simulated events where the total behavior is dictated by the real data, but the specific details of each event are still generated by the physical models. It is important to note that the underlying physical models used to generate these details are currently limited: they accurately describe the "Delta(1232)" resonance (a specific excited state of the nucleus) but do not yet include higher-energy resonances or deep inelastic scattering, which are necessary to describe the most energetic collisions.

When the researchers tested this new system against independent measurements from the CLAS and e4$\nu$ collaborations at the Thomas Jefferson National Accelerator Facility, the results were revealing. The system reproduced the total number of collisions in the energy range where the neural network was trained, confirming that the method works as intended, though it still fell short of the experimental data by about 15% at the quasi-elastic peak and 30% at the Delta peak. However, when they looked at the specific details of the debris, particularly in the region where heavier resonances appear, the simulation began to fall short of the experimental data. The system underestimated the number of events in these higher-energy regions, a limitation that stems from the fact that the underlying simulation models do not yet fully account for the most complex, high-energy interactions.

The study also highlighted a specific tension between different sets of experimental data. The new simulations matched one set of measurements from the Stanford Linear Accelerator Center almost perfectly, but they fell significantly below the results from the Jefferson Lab experiments at similar energies. This discrepancy suggests that there may be inconsistencies in the experimental data itself or that our understanding of the nuclear target needs further refinement. The researchers found that while their method could accurately predict the total number of collisions, it struggled to explain the full complexity of the particle debris in the semi-exclusive measurements, where only certain particles are detected.

Ultimately, this work provides a powerful new tool for the field. It demonstrates that it is possible to take a data-driven, machine-learning prediction of a total outcome and decompose it into the specific physical channels that generate it. This allows scientists to use the most accurate available data to guide their simulations while still retaining the detailed information about particle kinematics needed for complex experiments. While the current model has limitations, particularly in handling the most energetic collisions and the intricate interference between different quantum mechanical processes, it offers a clear path forward. By refining the underlying models and extending the method to other atomic targets, such as argon, this approach could significantly reduce the uncertainties that currently hinder our ability to measure the fundamental properties of the universe.