Improving neutrino energy estimation of charged-current interaction events with recurrent neural networks in MicroBooNE

The Big Picture

Imagine reconstructing a car crash from only the skid marks, broken glass, and displaced bumpers, never having witnessed the collision itself. That’s roughly the challenge neutrino physicists face. Neutrinos are nearly massless, electrically neutral particles that zip through matter at close to the speed of light, interacting so rarely that detecting even one requires massive, highly sensitive detectors. When they do interact, physicists must work backward from the debris to determine the original neutrino energy.

Getting that energy right is essential. Neutrino energy governs how particles oscillate, shifting between different flavors (electron, muon, tau) as they travel. These oscillations are one of the clearest windows into physics beyond the Standard Model, proving neutrinos have mass and potentially explaining why the universe contains matter rather than antimatter. None of that works without precise energy measurements.

The MicroBooNE collaboration, over 200 physicists across more than 40 institutions, has now shown that a recurrent neural network (RNN) outperforms conventional methods, with improvements validated on real experimental data from their detector in Illinois.

Key Insight: By feeding particle kinematic information into an RNN, MicroBooNE achieves reduced bias and improved energy resolution compared to the traditional approach of summing detected energy deposits, and the network holds up when tested against real data.

How It Works

The MicroBooNE detector is a liquid argon time projection chamber (LArTPC): a 170-ton tank of ultra-pure liquid argon that acts as both neutrino target and detector medium. When a neutrino strikes an argon nucleus, it produces a cascade of final-state particles, primarily a muon plus a spray of hadrons (protons, pions, and their kin). These particles ionize the argon, and the resulting electron clouds drift toward wire planes, producing 2D projections that get reconstructed into 3D particle tracks.

The traditional approach sounds simple enough: sum all the visible energy deposited in the detector. In practice, it runs into serious problems. Neutrons escape without leaving ionization trails. Nuclear binding energy gets absorbed by the argon nucleus. Detector gaps and inefficiencies pile on top.

The upshot: reconstructed energy consistently undershoots the true value, and even events at identical true energies scatter widely in the reconstruction.

The RNN takes a different tack. Instead of integrating raw energy deposits, it ingests the kinematic properties of individual reconstructed particles. The architecture handles variable-length sequences, which matters because the number of final-state particles differs from event to event. One interaction might produce a muon and a single proton; another might produce a muon, three protons, and several pions.

The input features include:

• Momentum magnitude and direction (polar and azimuthal angles) for each reconstructed particle
• Particle type (muon vs. hadron), identified using existing reconstruction tools
• A flag marking the primary muon candidate

The network processes each particle in turn, updating an internal hidden state that acts as a running summary, then outputs an energy estimate for the whole event. Training uses Monte Carlo (MC) simulated events that model realistic neutrino interactions, teaching the network to map reconstructed kinematics to true neutrino energy. After training, it runs inference on both held-out MC samples and real MicroBooNE data collected from 2016 to 2018.

Why It Matters

The validation is what sets this apart. Machine learning methods can latch onto features of simulated data that don’t exist in real detectors. A network that looks great in simulation might fall apart on actual data. The team guarded against this by running data-MC consistency tests, comparing RNN output distributions between real data and simulation across multiple kinematic regimes. The agreement held, confirming the network isn’t memorizing simulation artifacts.

The payoff shows up in physics reach. When tested against a neutrino oscillation analysis searching for sterile neutrino signatures (hypothetical particles that don’t interact via any known force), the RNN-based estimator improves sensitivity over the traditional approach. That improvement survives a full treatment of statistical and systematic uncertainties. In neutrino physics, those uncertainties are notoriously large, covering everything from interaction cross-sections to detector response modeling, so getting through that gauntlet intact is a meaningful result.

The RNN framework handles any set of reconstructed particles without assuming a fixed event topology. Other interaction channels, other LArTPC experiments, different detector technologies entirely: all are fair game. Upcoming experiments SBND and ICARUS (which share the same Fermilab neutrino beamline) and the next-generation DUNE detector are natural candidates.

One open question is how to handle systematic uncertainties directly within network training. Currently they are assessed after the fact. Future iterations that incorporate uncertainty-aware training could squeeze out further sensitivity gains.

Bottom Line: MicroBooNE’s RNN-based energy estimator outperforms the conventional visible-energy method in both bias and resolution, survives real-data validation, and improves sensitivity for neutrino oscillation searches. Deep learning can deliver for next-generation neutrino detectors.