⚛️ high-energy experiments

D $e^+e^-$ ffusion: Capturing the Beam-Beam Physics of $e^+e^-$ Collisions with Diffusion Models

The paper introduces D $e^+e^-$ ffusion, a permutation-equivariant diffusion model that serves as a fast and accurate surrogate for computationally expensive Monte Carlo simulations of incoherent pair creation backgrounds at high-luminosity $e^+e^-$ colliders like FCC-ee, generating events nearly four orders of magnitude faster than Geant4 while faithfully reproducing key kinematic and detector-level distributions.

Original authors: Antonio Chahine, Mariarosaria D'Alfonso, Jan Eysermans, Emmett Forrestel, Loukas Gouskos, Lindsey Gray, Katie Kudela, Haoyun Liu, Benedikt Maier, Dimitrios Ntounis, Christoph Paus, Umar Sohail Qureshi

Published 2026-07-22

📖 3 min read🧠 Deep dive

CC BY 4.0

Original authors: Antonio Chahine, Mariarosaria D'Alfonso, Jan Eysermans, Emmett Forrestel, Loukas Gouskos, Lindsey Gray, Katie Kudela, Haoyun Liu, Benedikt Maier, Dimitrios Ntounis, Christoph Paus, Umar Sohail Qureshi, Caterina Vernieri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, high-speed racetrack where tiny particles called electrons and positrons zoom around at nearly the speed of light. Scientists build massive machines, like the proposed Future Circular Collider (FCC-ee), to smash these particles together. The goal? To see what happens when they collide, hoping to uncover the deepest secrets of nature. But there's a catch: before the particles even hit each other, their intense electric fields act like a chaotic storm, creating a blizzard of extra particles called "beam-induced backgrounds." It's like trying to watch a delicate snowflake land on a table while a hurricane is blowing dust everywhere. To design detectors that can see the real action without getting blinded by the storm, scientists need to simulate billions of these background particles. However, doing this with current computer methods is like trying to count every grain of sand on a beach by picking them up one by one with tweezers—it takes too long and costs too much computing power. This is where a new kind of artificial intelligence steps in to save the day.

The paper introduces a clever new tool called De+e−ffusion, a type of AI designed to act as a "fast-forward" button for simulating these particle storms. Think of the traditional simulation method, called GuineaPig++, as a meticulous artist who paints every single drop of rain in a storm with perfect accuracy, but it takes them days to finish one painting. De+e−ffusion, on the other hand, is like a master impressionist who has studied millions of those paintings. It learns the patterns of how the rain falls, the wind blows, and the drops splash, and then it can instantly "dream up" a new, equally realistic storm in a fraction of a second.

The researchers trained this AI on a relatively small sample of about 60,000 simulated particle collisions. They taught it to understand not just the energy and speed of the particles, but also their specific "personality" based on how they were created. There are three main ways these background particles are born: a "Breit-Wheeler" dance (two real photons colliding), a "Bethe-Heitler" mix (one real and one virtual photon), and a "Landau-Lifshitz" tangle (two virtual photons). The AI learned to keep these three distinct styles separate, ensuring it didn't just blur them into a generic mess.

When the team tested their creation, the results were impressive. The AI-generated particles matched the slow, careful simulations almost perfectly in terms of their energy, speed, and where they appeared in space. To make sure the AI wasn't just faking it, they ran the generated particles through a full, detailed simulation of a real detector (the CLD vertex detector) and asked a sophisticated AI "judge" to tell the difference between the real and the fake. The judge struggled, achieving a score of only 0.553 (where 0.5 means it was guessing randomly). This suggests the AI's fake storms are nearly indistinguishable from the real thing.

The most exciting part is the speed. While the traditional method takes hours to simulate a single collision event, De+e−ffusion can do it in about 1.2 seconds on a powerful graphics card. That is a speedup of nearly 10,000 times. This doesn't mean the old method is wrong; it just means the new AI can do the heavy lifting so fast that scientists can finally run the massive simulations they need to design the next generation of particle detectors without waiting years for the computers to finish. It's a powerful new tool that turns a bottleneck into a breeze, paving the way for future discoveries in high-energy physics.

Technical Summary: De+e−ffusion

Problem Statement
High-luminosity $e^+e^-$ colliders, such as the Future Circular Collider (FCC-ee), face significant challenges from beam-induced backgrounds (BIB). The dominant component of these backgrounds is incoherent pair creation (IPC), where equivalent photon fields from colliding bunches produce $e^+e^-$ pairs via three QED subprocesses: real-photon Breit–Wheeler ( $\gamma\gamma \to e^+e^-$ ), mixed real-virtual Bethe–Heitler ( $\gamma\gamma^* \to e^+e^-$ ), and doubly-virtual Landau–Lifshitz ( $\gamma^*\gamma^* \to e^+e^-$ ).

The de facto standard for simulating these processes is GuineaPig++, a strong-field tracking particle-in-cell code. However, GuineaPig++ is computationally expensive, requiring several CPU-hours per bunch crossing. Comprehensive detector design studies (e.g., occupancy, tracking, radiation damage) require event samples orders of magnitude larger ( $O(10^5)$ bunch crossings) than what is feasible with current Monte Carlo (MC) generation. This computational bottleneck necessitates a fast, high-fidelity surrogate model.

Methodology
The authors introduce De+e−ffusion, a denoising diffusion probabilistic model (DDPM) designed as a permutation-equivariant, set-valued surrogate for IPC simulation. The model operates on a "particle cloud" representation of events, where each event is a set of particles with continuous kinematic/spatial features and discrete labels.

Key architectural and training innovations include:

Joint Continuous-Discrete Diffusion: The model diffuses a 7-dimensional continuous state vector (log-energy, unsquashed velocity components, and 3D position) alongside a discrete charge label ( $e^-$ $e^{-}$ or $e^+$ $e^{+}$ ).
- Velocity Unsquashing: To prevent unphysical superluminal samples, the velocity vector $\beta$ is mapped to $\mathbb{R}^3$ via an inverse hyperbolic tangent transformation before diffusion.
- Charge Encoding: The charge label is diffused using a discrete uniform-flip kernel (D3PM family).
Process Conditioning: A critical design choice is treating the QED subprocess label (Breit–Wheeler, Bethe–Heitler, Landau–Lifshitz) as a time-independent conditioning input that is never noised. This allows the model to learn the distinct longitudinal ( $z$ ) and kinematic structures of each subprocess, which would otherwise be averaged out in an unconditional model.
Loss Reweighting: Given the extreme class imbalance in the training data (Landau–Lifshitz dominates with $\sim85\%$ of particles, while Breit–Wheeler is $\sim1\%$ ), the denoising score-matching loss is reweighted using an inverse-frequency scheme ( $\alpha_w = 0.5$ ) to ensure rare channels are not washed out.
Auxiliary Objectives: The training loss includes:
- A cross-entropy term for the charge head to ensure accurate charge prediction.
- A charge conservation regularizer to enforce global event neutrality ( $e^+e^-$ pairs).
Architecture: The denoiser is a transformer encoder (3 layers, 2 heads, hidden dimension 512) with shared trunk and separate heads for noise prediction (continuous) and charge logits (discrete).

Key Contributions

Fast Simulation Surrogate: De+e−ffusion provides a generative model that faithfully reproduces the marginal and joint distributions of IPC events.
Process-Aware Modeling: By conditioning on the QED subprocess, the model successfully resolves the multi-modal nature of the longitudinal production profiles, a feature lost in unconditional generative approaches.
Detector-Level Validation: The authors perform a rigorous two-sample test by propagating both GuineaPig++ and De+e−ffusion events through a full Geant4 simulation of the CLD vertex detector. A transformer-based classifier is trained to distinguish between the two.
Performance Benchmarking: The paper benchmarks the generation speed against GuineaPig++ on both CPU and GPU hardware.

Results

Particle-Level Fidelity: The model reproduces one-dimensional marginals of energy, momentum, transverse momentum ( $p_T$ ), velocity components, and spatial coordinates ( $x, y, z$ ) with percent-level agreement to GuineaPig++ across the bulk of distributions, including heavy tails.
Subprocess Resolution: The longitudinal $z$ distributions for all three subprocesses (Breit–Wheeler, Bethe–Heitler, Landau–Lifshitz) are accurately reproduced, capturing the distinct production vertices (central for BW, forward for LL).
Detector-Level Agreement: When passed through Geant4, the generated hits show near-closure in spatial ( $x, y, z$ ) and energy deposition ( $E_{dep}$ ) distributions.
Classifier Test: A transformer-based classifier trained to distinguish GuineaPig++ hits from De+e−ffusion hits achieves an Area Under the ROC Curve (AUC) of 0.553 $\pm$ 0.016. This value is only modestly above the random guessing baseline (0.5) and falls within the noise floor established by training a classifier on two halves of the same GuineaPig++ sample (AUC $\approx$ 0.522). This indicates the generated events are statistically indistinguishable from the reference at the detector level.
Speedup: De+e−ffusion generates events nearly four orders of magnitude faster than GuineaPig++ when running on a GPU with batched sampling (approx. 1.2 seconds per event vs. 15,000 seconds). Even on CPU without batching, it offers a $5\times$ speedup.

Significance and Claims
The paper claims that De+e−ffusion successfully alleviates the computational bottleneck of IPC simulation for FCC-ee design studies. It provides a high-fidelity, fast-simulation surrogate that enables the generation of event samples ( $O(10^5)$ bunch crossings) required for comprehensive detector and machine-detector interface studies, which are currently intractable with standard MC generators.

The authors note that while the current model is conditioned on fixed machine parameters (FCC-ee Z-pole), future work aims to condition on machine parameters (bunch sizes, intensities, crossing angles) to enable rapid parameter space scans. They also acknowledge that while WarpX is emerging as a GPU-accelerated alternative to GuineaPig++, De+e−ffusion offers a distinct generative approach that, once trained, provides immediate, massive speedups for specific design configurations. The work is presented as a step toward a detector-agnostic, high-fidelity simulation framework for beam-background studies.

Technical Summary: De+e−ffusion

More like this