[Paper Review] PIPPIN: Generating variable length full events from partons

G. Quétant, J. A. Raine+3 more|arXiv (Cornell University)|Jun 18, 2024

Particle physics theoretical and experimental studies4 citations

TL;DR

PIPPIN introduces a novel machine learning framework that directly generates variable-length detector-level event point clouds from parton-level inputs using transformers, score-based models, and normalizing flows. It achieves state-of-the-art performance in simulating top quark pair events at the LHC, accurately capturing multiplicity, kinematics, and topological correlations without autoregressive generation or fixed output lengths.

ABSTRACT

This paper presents a novel approach for directly generating full events at detector-level from parton-level information, leveraging cutting-edge machine learning techniques. To address the challenge of multiplicity variations between parton and reconstructed object spaces, we employ transformers, score-based models and normalizing flows. Our method tackles the inherent complexities of the stochastic transition between these two spaces and achieves remarkably accurate results. The combination of innovative techniques and the achieved accuracy demonstrates the potential of our approach in advancing the field and opens avenues for further exploration. This research contributes to the ongoing efforts in high-energy physics and generative modelling, providing a promising direction for enhanced precision in fast detector simulation.

Motivation & Objective

• To develop a fast, accurate, and end-to-end generative model for high-energy physics event simulation that bypasses traditional step-by-step simulation pipelines.
• To address the challenge of variable particle multiplicity between parton and reconstructed object spaces in detector-level simulation.
• To enable conditional generation of full events from parton inputs while preserving complex kinematic correlations and topological structures.
• To explore the feasibility of direct, non-autoregressive generation of reconstructed objects from partons using modern deep generative architectures.
• To demonstrate the model’s generalization capability across multiple decay channels and its potential for both simulation and unfolding tasks.

Proposed method

• The model employs two Transformer encoders to process input partons and predict the number of reconstructed objects via a multiplicity predictor.
• A PIP-Droid generator, based on score-based generative modeling, conditionally generates detector-level point clouds of variable length from the encoded parton information.
• The architecture integrates normalizing flows to model complex, high-dimensional distributions of reconstructed particles while maintaining permutation invariance for unordered particle sets.
• The model is trained end-to-end using a conditional diffusion process, enabling stochastic generation of multiple event instances from the same parton input.
• Permutation invariance ensures robustness to particle ordering, and the architecture avoids truncation or autoregressive generation by predicting output length directly.
• The framework supports conditional control over event properties, such as fixed numbers of jets or leptons, enabling targeted simulation and phase space filtering.

Figure 1: Diagram of the global architecture and the training processes of the PIPPIN model. It is made of two Transformer Encoders, which encode the partons, a Multiplicity Predictor, which predicts the number of reconstructed objects, and a PIP-Droid Generator, which conditionally generate these r

Experimental results

Research questions

• RQ1 Can a single deep generative model accurately simulate full detector-level events from parton-level inputs without step-by-step simulation?
• RQ2 How well can a model with variable-length output generation preserve complex kinematic correlations and topological structures in multi-channel decay processes?
• RQ3 To what extent can a single model trained on inclusive top quark pair events generalize to specific decay channels without channel-specific fine-tuning?
• RQ4 How does the performance of the proposed method compare to existing fast simulators and full Monte Carlo simulations in terms of accuracy and efficiency?
• RQ5 Can the model be adapted for the reverse problem of unfolding reconstructed objects back to partons, and what would be the key challenges in such a setup?

Key findings

• PIPPIN significantly outperforms both Turbo-Sim and a baseline diffusion model in reconstructing key observables such as leading jet energy and t-tbar invariant mass, with results closely matching the original Monte Carlo simulation.
• The model achieves high fidelity in capturing multiplicity and kinematic distributions across multiple decay channels, including all-hadronic, semi-leptonic, and fully leptonic final states, without channel-specific training.
• The marginal distributions of reconstructed jet energy and t-tbar system mass from PIPPIN closely follow the reference Monte Carlo distributions, with small deviations within statistical uncertainties.
• PIPPIN maintains strong performance even when evaluated on a restricted subset of events (e.g., with leptons), outperforming models trained specifically on that subset, indicating superior generalization.
• The model’s ability to generate variable-length outputs directly, without truncation or autoregressive generation, enables efficient and accurate simulation across diverse event topologies.
• The conditional generation capability allows for controlled simulation, such as fixing the number of jets or leptons, enabling targeted exploration of phase space.

Figure 12: The Transformer Encoder architecture details.