PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning and LHC Event Generation

Anja Butter, Tilman Plehn, Steffen Schumann, Simon Badger, Sascha Caron, Kyle Cranmer, Francesco Armando Di Bello, Etienne Dreyer, Stefano Forte, Sanmay Ganguly, Dorival Gonçalves, Eilam Gross, Theo Heimel, Gudrun Heinrich, Lukas Heinrich, Alexander Held, Stefan Höche, Jessica N. Howard, Philip Ilten, Joshua Isaacson, Timo Janßen, Stephen Jones, Marumi Kado, Michael Kagan, Gregor Kasieczka, Felix Kling, Sabine Kraml, Claudius Krause, Frank Krauss, Kevin Kröninger, Rahool Kumar Barman, Michel Luchmann, Vitaly Magerya, Daniel Maitre, Bogdan Malaescu, Fabio Maltoni, Till Martini, Olivier Mattelaer, Benjamin Nachman, Sebastian Pitz, Juan Rojo, Matthew Schwartz, David Shih, Frank Siegert, Roy Stegeman, Bob Stienen, Jesse Thaler, Rob Verheyen, Daniel Whiteson, Ramon Winterhalder, Jure Zupan

arXiv:2203.07460v2hep-phhep-ex

TL;DR

LHC simulations connect detector data with fundamental theory, but event generators face increasingly demanding precision and complexity challenges. This review surveys machine-learning approaches to event generation and inverse simulation, describing applications across the simulation chain and the challenges that remain.

  • Problem

    Event generators must address complex high-multiplicity final states and phase spaces, while parton showers use a semiclassical approximation beyond the small-angle regime where it is justified.

    Source: §2.1 · p. 5, §2.4 · p. 9

  • Method

    The review surveys machine-learning applications to event generation and inference, including learned phase-space distributions, parton-level unfolding, and sharing generative models.

    Source: §1 · p. 5, §2.1 · p. 6, §4.3 · p. 19

  • Results

    The review describes machine learning as offering approaches to event-generator challenges and inverse simulation, while noting that unfolding methods have distinct strengths and weaknesses.

    Source: §1 · p. 5, §4.2 · p. 18

  • Takeaways & Limitations

    Parton-level unfolding can produce probability distributions of configurations for defining observables, and generative networks can provide an efficient alternative for sharing simulated data.

    Source: §4.3 · p. 19, §5 · p. 23

    Generalization remains challenging when information is scarce or absent, including extrapolation to kinematic regions without data.

    Source: §2.5 · p. 11

Abstract

from arXiv · show

First-principle simulations are at the heart of the high-energy physics research program. They link the vast data output of multi-purpose detectors with fundamental theory predictions and interpretation. This review illustrates a wide range of applications of modern machine learning to event generation and simulation-based inference, including conceptional developments driven by the specific requirements of particle physics. New ideas and tools developed at the interface of particle physics and machine learning will improve the speed and precision of forward simulations, handle the complexity of collision data, and enhance inference as an inverse simulation problem.

1 Introduction

First-principle simulations link collider data to fundamental theory, and precision and speed are both essential. The introduction sets out how machine learning may help meet rising simulation challenges and support different kinds of LHC measurements.

1 Introduction

  • The simulation chain connects theory predictions to particle-level events and detector simulation, and should rely on first-principles physics.The introduction frames precision and speed as equally essential.Source: §1 · p. 4
  • The introduction distinguishes theory-independent experimental measurements, measurements interpreted through fundamental theory, and measurements comparing data with theory to probe fundamental physics.Machine learning may help analyze and calibrate data, extract parton densities, improve Monte Carlo simulations, and search for deviations or measure theory parameters.Source: §1 · p. 4
  • The projected increase from 139 fb−1 in Run 2 to 4 ab−1 at the HL-LHC suggests that theory predictions must match experimental uncertainties at and below the percent level.The passage also notes increased rates, more exotic kinematic regions, and more complex signatures as challenges for established event generators.Source: §1 · pp. 4–5
  • The review discusses machine-learning approaches to event-generator challenges, end-to-end ML generators, and inverse simulation, including simulation-based inference.Figure 1 depicts the simulation chain; the review discusses its forward direction in Sections 2–3 and inverse simulation in Section 4.Source: Figure 1 · p. 4, §1 · p. 5

2 Machine Learning in event generators

Machine learning can support multiple stages of event generation, from phase-space sampling and amplitude calculations to parton showers, PDFs, and fragmentation functions. The review describes both approaches that integrate ML into existing simulation components and methods that learn event distributions or physics inputs.

2.1 Phase space sampling

  • Trainable phase-space maps and channel weights can adapt sampling to complex integrands, while direct generative models learn event distributions from weighted or unweighted samples.For high-multiplicity processes, reported gains over established methods can fall below unity; differentiable matrix-element gradients are also being explored to improve model training.Source: §2.1 · p. 6

2.2 Scattering Amplitudes

  • Neural-network surrogates and physics-informed architectures address amplitude evaluation and precision, with factorization-aware parametrization reaching per-mille precision for tree-level e+e−→5 jets.The review notes that this precision does not translate into a clear improvement of higher-order LHC predictions; a single model spanning varied cuts and settings could potentially improve overall performance further.Source: §2.2 · pp. 7–8, Figure 3 · p. 8

2.3 Loop integrals

  • Normalizing flows can optimize Feynman-parameter integration domains and improve contour choices when standard methods produce sub-optimal contours for complicated integrals or specific phase-space regions.The network must satisfy contour boundary conditions associated with constraints such as the Landau equations and Cauchy’s theorem.Source: §2.3 · pp. 8–9, Figure 4 · p. 9

2.4 Parton shower

  • Parton-shower learning can scaffold a model over the semiclassical branching framework or use set- and graph-based architectures to encode permutation symmetry.An alternative learns shower-specific splitting kernels from QCD predictions and data, but remains limited by the applicability of the simple-splittings picture.Source: §2.4 · pp. 9–10

2.5 Parton distribution functions

  • PDF determinations use Monte Carlo ensembles of machine-learning models to represent probability distributions over possible PDFs, with modern workflows adding tools such as cross-validation and hyperoptimization.A continuing challenge is avoiding overfitting and generalizing to regions with scarce or no data; judging overfitting still relies partly on the fitter’s accumulated PDF knowledge.Source: §2.5 · pp. 10–11

2.6 Fragmentation functions

  • Feed-forward neural networks provide data-driven interpolants for fragmentation functions, and replica-trained networks use their spread to estimate and propagate input-data uncertainties.Combining these functions with higher-order perturbative QCD calculations provides predictions for hard-scattering processes with identified final-state hadrons.Source: §2.6 · p. 12

3 End-to-end ML-generators

End-to-end ML generators directly produce events, complementing machine-learning improvements to conventional generator modules. The section compares generative architectures and describes methods to control their phase-space precision and quantify uncertainty.

3 End-to-end ML-generators

  • End-to-end ML generators directly generate events and can support phase-space studies, inverted simulations, phenomenological datasets, and efficient event-sample delivery.They can also train on data combined with simulations, manipulate samples, or post-process Monte Carlo data, including to unweight events.Source: §3 · p. 12
  • The main ML-generator architectures are GANs, VAEs, optimal-transport probabilistic autoencoders, normalizing flows, and invertible-network variants.Neural-network interpolation and implicit bias can amplify statistics-limited training data.Source: §3 · p. 13

3.1 Fast generative networks

  • VAEs map latent space to phase space with a decoder and approximate the inverse with an encoder, but ELBO optimization and balancing reconstruction against KL divergence can limit performance.A B-VAE uses β ≪ 1 and, after optimization, replaces the Gaussian latent distribution with a buffer derived from training events.Source: §3.1 · p. 13
  • GANs reproduce LHC event phase-space densities, but may lack precision for sharp parton-level features and can require a targeted MMD loss.The central challenge is achieving precision in the underlying density while finding a Nash equilibrium.Source: §3.1 · pp. 13–14
  • Normalizing flows trade some flexibility for tractable likelihood evaluation, can train on negative-weight events, and offer tools for performance control and uncertainty quantification.They use bijective transformations with tractable Jacobians; their invertibility also supports anomaly detection and density estimation.Source: §3.1 · p. 14
  • Optimal-transport probabilistic autoencoders learn theory-to-data and data-to-theory mappings without paired event samples.Their latent space represents parton-level information, and their encoder and decoder define conditional mappings that marginalize to theory- and data-space priors.Source: §3.1 · p. 14
  • OTUS learned a qualitatively simulator-like mapping despite training only on marginal matching, even recovering behavior around invariant masses withheld during training.The authors suggest physically meaningful mappings may emerge even when relations are not included in training; supplying known relations as inductive biases may improve performance.Source: §3.1 · pp. 14–15

3.2 Control and precision

  • Generative networks need control over relevant phase-space features and estimates of how precisely they learned the underlying density.A discriminator can identify deviations and support reweighting or joint training; Bayesian weight distributions provide training-related uncertainty, which increases in kinematic tails with less training data.Source: §3.2 · pp. 15–16

4 Inverse simulations and inference

Machine-learning methods make inverse simulation and inference more tractable across particle reconstruction, detector and parton-level unfolding, and parameter estimation. They offer high-dimensional, unbinned approaches while retaining correlations and uncertainty information.

4.1 Particle reconstruction

  • Machine-learning reconstruction improves neutral-energy resolution by a factor of two to four over traditional particle flow in a simplified calorimeter study.Graph neural networks also handle realistic pileup environments by using spatial relations among irregular, sparse detector inputs.Source: §4.1 · p. 17

4.2 Detector unfolding

  • Classifier-based reweighting and density-based generative methods enable high-dimensional, binning-independent unfolding, with post-hoc choice of observables and bins.The approaches have distinct strengths and weaknesses, and combining them requires further research and development to achieve precision.Source: §4.2 · p. 18

4.3 Unfolding to parton level

  • Conditional generative networks can infer parton-level configurations from detector events, enabling observables in their original reference frames without binning.The learned probability distributions account for QCD radiation and decays under the forward-simulation model, including its assumed hard process and number of hard-scattering jets.Source: §4.3 · pp. 18–19
  • A cINN resolves W-decay jet combinatorics, producing identical reconstructed W-widths across jet multiplicities.The network learns the relation between up to four jets and the two partonic W-decay quarks using correlations and forward-simulation truth.Source: §4.3 · p. 19

4.4 MadMiner

  • MadMiner trains neural networks from joint likelihood ratios to approximate likelihood ratios for detector-level data and support parameter confidence limits.The simulation-based model can describe high-dimensional data, while score-based observables are sufficient statistics for parameter points close enough to the reference point.Source: §4.4 · pp. 20–21

4.5 Matrix element method

  • Learned transfer functions and conditional flows can support event-by-event matrix-element likelihoods that use full correlated kinematics.This approach targets observables including the top mass, Higgs self-coupling, and CP-violating phases; differing parton- and detector-level phase spaces and difficult numerical convolutions motivate the learned treatment.Source: §4.5 · pp. 21–22

5 Synergies, transparency and reproducibility

Sharing machine-learning models and data can support reuse, reproducibility, and synergies across collider analyses. Generative models offer an additional route to sharing simulated data, but practical reuse still faces data and documentation challenges.

Model sharing

  • Sharing neural networks can help researchers examine and build on results, enable synergies across use cases, and facilitate reproducibility.Models may encode scattering matrix elements or probability models used in simulation-based inference.Source: §5 · pp. 22–23

Generative models

  • Generative networks can be shared through architecture and weights alone and can also serve as an efficient way to share simulated data.The passage notes an opportunity for phenomenological studies based on the DELPHES detector simulation.Source: §5 · p. 23

Reuse challenges

  • Model reuse is not yet standard, and missing input-variable documentation complicates publication for future use.Sharing also requires suitable input data; open data and software containers are identified as useful parts of the effort.Source: §5 · p. 23

6 Outlook

Machine learning can improve every aspect of modular event generation, while generative networks and simulation-based inference extend the review’s outlook to fast simulation and inverse problems. These tools address LHC needs and may also contribute to broader machine-learning research.

6 Outlook

  • Machine learning is a transformative addition to particle physics’ numerical toolbox, though it will not solve all LHC simulation and analysis challenges.The review expects particle physics to eventually contribute to broader machine-learning research, given the quality of its datasets, simulations, and simulation-based analysis methods.Source: §6 · p. 23
  • Machine learning can improve phase-space sampling, scattering amplitudes, loop integrals, parton showers, parton densities, and fragmentation, all of which will be key for HL-LHC needs.Some applications are established approaches, while others are rapidly developing and entering standard generators.Source: §6 · pp. 23–24
  • Generative networks at parton level and fast ML-detector simulations provide testing grounds for phase-space generators, precision networks, and inverted simulations.LHC requirements motivate precision control and full uncertainty ranges; these networks can also define, produce, and encode datasets for phenomenological studies and compress data entering experimental analyses.Source: §6 · p. 24
  • Simulation-based inference can use classification, conditional generative networks, or learned likelihood ratios to predict model-parameter distributions from observed LHC events.The review describes potential applications from particle identification and detector unfolding to event-wise matrix-element analyses at the HL-LHC.Source: §6 · p. 24

Related papers

Loading 2203.07460v2…