Abstract
Neural networks for LHC physics have to be accurate, reliable, and controlled. Using neural surrogates for the prediction of loop amplitudes as a use case, we first show how activation functions are systematically tested with Kolmogorov-Arnold Networks. Then, we train neural surrogates to simultaneously predict the target amplitude and an uncertainty for the prediction. We disentangle systematic uncertainties, learned by a well-defined likelihood loss, from statistical uncertainties, which require the introduction of Bayesian neural networks or repulsive ensembles. We test the coverage of the learned uncertainties using pull distributions to quantify the calibration of cutting-edge neural surrogates.
Introduction
In the era of high-precision particle physics, the computational demand for simulating Large Hadron Collider (LHC) events is immense. A significant bottleneck is the calculation of "loop amplitudes," which describe complex particle interactions. To address this, physicists employ neural network surrogates—fast approximations of these rigorous calculations. However, for these surrogates to be viable in precision studies, they must meet two critical criteria: extreme accuracy (errors negligible compared to theoretical uncertainties) and trustworthy uncertainty quantification (calibrated error estimates).
Enhancing Expressivity: KANs and Symmetry
A primary focus of this work is optimizing the architecture for maximum precision. The authors investigate Kolmogorov-Arnold Networks (KANs) and GroupKANs, which utilize learnable activation functions instead of fixed ones like ReLU. While KANs offer a systematic way to test activation functions, the study finds that standard networks with GELU activations perform comparably well for this specific task.
More significant gains were achieved by incorporating physical symmetries. Moving from standard Multi-Layer Perceptrons (MLPs) to architectures like Deep Sets Invariants (DSI) and Lorentz Geometric Algebra Transformers (L-GATr)—which are designed to respect the underlying symmetries of particle physics—reduced relative errors from $10^{-2}$ to a remarkable $10^{-5}$.
Disentangling Uncertainties
Reliability in neural surrogates requires distinguishing between different sources of error:
- Systematic Uncertainties: Arising from model limitations or noise in the training data. These are captured using a heteroscedastic loss, where the network predicts both a mean and a variance ($\sigma$).
- Statistical Uncertainties: Arising from finite training data. These are modeled using Bayesian Neural Networks (BNNs), which treat weights as distributions, and Repulsive Ensembles (REs), which use a mathematical kernel to keep ensemble members diverse in parameter space.
To ensure these uncertainties are "honest," the authors utilize pull distributions. By checking if the residuals (prediction minus truth) normalized by the predicted uncertainty follow a standard unit Gaussian, they can quantify the calibration of the model.
Results and Calibration
The benchmark task involved predicting squared loop-induced matrix elements for the process $g g \to \gamma\gamma g$ using a dataset of 1.1 million phase-space points. The key findings include:
- High Precision: DSI and L-GATr models achieved a median relative accuracy of $10^{-5}$.
- Noise Recovery: In tests with artificial Gaussian noise, the models successfully learned to predict the injected noise level as a systematic uncertainty.
- Calibration Success: BNNs and calibrated Repulsive Ensembles produced nearly perfect unit Gaussian pull distributions, indicating that their self-reported error bars are statistically reliable. Ensembles initially showed under-confidence regarding systematic errors, which was corrected through a temperature calibration step.
Conclusion
This research demonstrates that neural surrogates can achieve the sub-percent precision required for LHC physics while providing robust, calibrated uncertainty estimates. By combining symmetry-aware architectures with Bayesian frameworks, the authors provide a path toward more reliable and interpretable machine learning tools in high-energy physics. Future work will continue to explore the interplay between model expressivity and data limits as the field pushes toward even higher precision benchmarks.
Summary updated