Abstract
Generative modeling of high-energy physics data requires handling hybrid datasets containing both continuous kinematic variables and discrete quantum properties. We present a Multimodal Generative Flow (MMF) framework that integrates continuous flow-matching with discrete Markov jump processes within a unified transformer architecture. Evaluated on the ASPENOPENJETS dataset, the MMF model demonstrates superior fidelity in capturing cross-modal correlations, significantly improving the reproduction of complex observables like jet charge and substructure variables compared to existing baselines.
Introduction
In high-energy physics, "jets" are collimated sprays of particles produced in proton-proton collisions at the Large Hadron Collider (LHC). These are typically modeled as particle clouds where each particle possesses continuous features (momentum $p_T$, position $\eta, \phi$) and discrete categorical identities (flavor and charge).
Traditional generative models often struggle with this hybrid nature, relying on de-quantization or separate modeling that fails to capture physical correlations between a particle's identity and its kinematics. This paper introduces a way to treat both types of data as part of a single, unified dynamical flow, ensuring that the evolution of a particle's position is mathematically synchronized with the evolution of its identity.
Methods
The MMF framework introduces a unified probabilistic path for hybrid spaces by integrating two mathematical bridges:
- Continuous Flow-Matching: Used for particle kinematics, learning a velocity field that pushes noise toward the target data distribution.
- Continuous-Time Markov Jump Bridge: A multi-state generalization of the random telegraph process used for discrete flavor tokens.
The architecture, termed the Multimodal ParticleFormer, uses a permutation-equivariant transformer with mode-specific encoders that feed into a fused encoder. This allows the model to learn the intricate relationships between a particle's type and its motion. Training utilizes an uncertainty-weighted loss function to balance Mean Squared Error and Cross-Entropy automatically, while sampling is performed using $\tau$-leaping to efficiently simulate discrete state transitions.
Results
The model was trained on 1.25 million jets from the ASPENOPENJETS (AOJ) dataset, derived from 13 TeV CMS data. MMF significantly outperformed the EPiC-FM baseline in reproducing jet substructure variables such as N-subjettiness ratios ($\tau_{21}, \tau_{32}$), which describe the internal shape of the jet.
Notably, MMF achieved a Wasserstein distance for Jet Charge of $1.4 \times 10^{-3}$, nearly an order of magnitude better than the baseline ($9.5 \times 10^{-3}$). This proves its superior ability to link kinematics with discrete quantum numbers. The model also demonstrated higher spatial accuracy in centering jets within the detector coordinate plane.
Conclusion
The MMF approach provides a robust solution for generating high-fidelity multimodal particle clouds. While the model excels at dominant particle types, performance on extremely rare classes like leptons remains a bottleneck due to data imbalance. Future work will focus on refining sampling noise issues and exploring cross-attention mechanisms to further enhance cross-modal learning.
Summary updated