Observable Optimization for Precision Theory: Machine Learning Energy Correlators

The Big Picture

Imagine measuring the exact weight of a specific ingredient in a complex dish by analyzing just the smell. You can’t sample everything at once; you have to choose which scent to focus on. Pick the wrong one and you’ll get a muddy signal buried in noise. Pick the right one and you might nail it precisely. Physicists at particle colliders face exactly this problem, except instead of smells, they’re choosing which mathematical “shape” of particle data to examine.

Every time protons smash together at the Large Hadron Collider, they spray out hundreds of particles in a chaotic cascade. The raw data is immense. Tracking the energy and direction of every particle in every collision produces millions of numbers per event. To extract physics from this flood, scientists compress that information into a single number called an observable.

For simple sorting tasks, you can hand the job to a neural network and let it do whatever it wants. But for precision measurements, the kind that pin down fundamental constants to many decimal places, your observable must be something you can calculate from first principles. Neural networks are not that. They’re black boxes, and theoretical physicists can’t derive predictions for what a black box will output.

So the most powerful tools for finding useful signals are exactly the tools you can’t use for the most important measurements. A team of researchers, with support from IAIFI, found a way to thread this needle. They use machine learning to find the best observable, then step away, leaving behind only a clean, calculable quantity ready for comparison with data.

Machine learning can scout the space of possible precision observables and identify the best one, then exit the picture entirely, leaving a human-interpretable, theoretically computable result.

How It Works

The researchers focused on a specific class of observables called energy correlators, which measure how particle energies are distributed across different angular separations. The energy 3-point correlator (E3C) captures the pattern of energy flow by looking at the angles between three particles simultaneously.

The catch: a three-particle triangle has many possible shapes, forming a two-parameter family. Scientists typically collapse it to a single number by “marginalizing,” meaning they sum over all triangle shapes satisfying some constraint. But which constraint? Which triangle shape carries the most information about the top quark mass?

The approach works in two stages:

1. Learn the distribution. Using Monte Carlo simulations, the team produced large datasets of E3C data at different top quark mass values. They trained two types of density estimators: a dense neural network (DNN) with a weighting scheme that emphasizes physics-relevant features, and a normalizing flow, a type of neural network that explicitly models the probability of any given data point.

2. Search for the optimal observable. With a learned surrogate model in hand, they deployed neural ratio estimation (NRE), a simulation-based inference technique. NRE trains a classifier to distinguish samples generated at the true parameter value from samples with mismatched values. The likelihood ratio it learns directly measures the statistical power of any proposed observable. By sweeping through different triangle shape constraints, parameterized by the side ratios of isosceles triangles, they quantified how sharply each shape constrains the top quark mass.

The search is efficient because the neural surrogate is fast to evaluate. That makes it practical to scan a wide space of candidates without rerunning expensive simulations for each one.

Here’s the payoff. After scanning the full space of isosceles triangle marginalizations, they found that right triangles (side ratios of 1:1:√2) give the sharpest sensitivity to the top quark mass. That’s a clean, geometric result. A right triangle on the angular sphere is a precise definition that any theorist can write down and any experimentalist can implement.

Once this shape is identified, the machine learning infrastructure becomes disposable. Future measurements can simply report the E3C marginalized over right triangles, compute theoretical predictions using perturbative QCD, and compare directly to data. No neural networks remain in the final analysis chain.

As a consistency check, the team verified that marginals computed from the learned distributions matched those computed directly from simulation. Both the DNN and normalizing flow approaches gave consistent results.

Why It Matters

Precision particle physics has a bottleneck. The field has adopted machine learning for tasks where interpretability isn’t required: jet tagging, anomaly detection, fast simulation. But precision measurements of Standard Model parameters demand a different standard. Theoretical predictions must be computed order-by-order in perturbation theory, and that only works for well-defined mathematical quantities, not neural network outputs.

By framing observable optimization as a machine learning problem that terminates in a classical result, the authors open a new strategy. The same methodology could optimize observables for measuring the strong coupling constant, probing new physics signals, or extracting other fundamental parameters from LHC data. Any precision measurement that currently settles for a “good enough” observable could be revisited with this search technique.

Simulation-based inference has come a long way in recent years. Rather than approximating likelihood functions analytically, which is often impossible for complex collider processes, these methods extract statistical power directly from simulated data. That sidesteps years of difficult theoretical approximation.

By using ML to navigate the space of theoretically calculable observables and identify right-triangle energy correlators as optimal for top quark mass measurements, this work makes the case that machine learning and precision physics aren’t opposites. ML does the scouting; theory does the measurement.

---