Machine Learning in Nuclear Physics

The Big Picture

Imagine trying to solve a jigsaw puzzle with a trillion pieces, where every piece connects to every other piece in ways governed by quantum mechanics. That is, roughly speaking, what nuclear physicists face every day.

The atomic nucleus, a tiny bundle of protons and neutrons crammed into a space millions of times smaller than a single atom, is one of the most complex physical systems in the universe. Predicting how nuclei behave, how they smash together at high energies, or how they decay requires calculations of immense difficulty. Classical computational methods strain under the load.

A new review by researchers from Jefferson Lab, MIT, Michigan State, Lawrence Berkeley, and more than a dozen other institutions surveys how machine learning is reshaping this century-old field. The paper covers everything from detector design to taming mathematical obstacles that have blocked certain quantum calculations for decades. This is not a preview of future breakthroughs. The change has already happened.

Key Insight: Machine learning is not just speeding up calculations in nuclear physics. It is making previously intractable problems solvable, from predicting nuclear structure properties to reconstructing particle collisions in real time.

How It Works

The authors organize ML applications into four major areas: nuclear theory, experimental methods, accelerator science, and nuclear data. Each uses different architectures for fundamentally different problems.

Start with nuclear theory. Low-energy nuclear theory, the study of how forces between protons and neutrons determine nuclear structure, has long been limited by quantum-mechanical complexity. Even moderate nuclei contain dozens of interacting particles, and exact solutions are computationally forbidden. ML now provides surrogate models that predict nuclear masses, binding energies, and decay rates orders of magnitude faster than traditional methods.

The toolbox includes Gaussian processes (statistical techniques that express predictions as probability distributions rather than single values), Bayesian neural networks (which track uncertainty alongside predictions), and deep feedforward networks. These methods interpolate and extrapolate nuclear properties across the chart of nuclides, the periodic-table-like map of all known atomic nuclei.

Lattice QCD, the leading computational approach to the theory of quarks and gluons, is where ML tackles two of the hardest outstanding problems. The sign problem arises because the mathematical weight assigned to each possible field configuration can become complex rather than positive, causing catastrophic cancellations when summing millions of contributions. ML-based flow models generate configurations that sidestep the worst of these cancellations.

Normalizing flows and other generative models learn to sample from complex probability distributions. They can produce independent lattice field configurations far more efficiently than traditional Markov-chain Monte Carlo methods, which generate configurations one laborious step at a time. Flow-based approaches could cut the cost of a single lattice calculation by orders of magnitude.

On the experimental side, the challenges are different but equally brutal:

• Charged particle tracking in detectors produces millions of hits per second; ML-based graph neural networks reconstruct particle trajectories in real time with accuracy rivaling hand-tuned algorithms
• Calorimetry uses convolutional neural networks to extract energies and positions from pixel-level detector readout
• Particle identification and event classification use boosted decision trees and deep networks to distinguish signal from noise in flooded data environments
• Streaming detector readout poses a data-volume problem that ML-based trigger systems address by compressing and filtering data on the fly, before it can be written to disk

Accelerator science brings its own headaches. Modern accelerators are enormously complex machines with thousands of coupled parameters. Surrogate models trained on simulation data replace expensive physics-based models during real-time optimization, enabling feedback loops fast enough to correct beam instabilities as they emerge.

Accelerator systems are also non-stationary: components age, conditions drift, and a model trained yesterday may fail today. Adaptive ML techniques are built to handle exactly this kind of moving target.

Why It Matters

The review’s real contribution isn’t any single technique. It’s the convergence it documents.

Nuclear physics covers roughly fifteen orders of magnitude in energy scale, from binding energies of a few MeV to quark-gluon plasma temperatures above a trillion Kelvin. Experiments at CERN, Jefferson Lab, and RHIC generate petabytes of data annually. The theoretical models that interpret this data are some of the most computationally expensive in science. Machine learning is not a convenience here. It is becoming infrastructure.

The influence runs both ways. Techniques developed for nuclear problems (uncertainty quantification in neural networks, equivariant architectures that respect physical symmetries, flow-based samplers for quantum field theory) are feeding back into ML research. Nuclear physicists are becoming ML innovators. Tools built for hadron spectroscopy (classifying subatomic particles by quark content) or accelerator control are shaping how the broader community thinks about physics-informed machine learning.

Open questions remain sharp. How do you rigorously propagate uncertainties through deep neural networks used as theory emulators? Can generative models ever fully solve the sign problem, or only alleviate it? How do you ensure that ML-based event reconstruction does not introduce subtle biases that corrupt precision measurements? The review does not paper over these tensions. It is as direct about the open methodological challenges as it is about the successes.

Bottom Line: Machine learning has already reshaped nuclear physics across theory, experiment, and infrastructure. This review maps that transformation, revealing a field not waiting for AI but actively co-developing it.