Fast Point Cloud Generation with Diffusion Models in High Energy Physics

Vinicius Mikuni,Benjamin Nachman,Mariel Pettee

DOI: https://doi.org/10.1103/PhysRevD.108.036025

2023-07-18

Abstract:Many particle physics datasets like those generated at colliders are described by continuous coordinates (in contrast to grid points like in an image), respect a number of symmetries (like permutation invariance), and have a stochastic dimensionality. For this reason, standard deep generative models that produce images or at least a fixed set of features are limiting. We introduce a new neural network simulation based on a diffusion model that addresses these limitations named Fast Point Cloud Diffusion (FPCD). We show that our approach can reproduce the complex properties of hadronic jets from proton-proton collisions with competitive precision to other recently proposed models. Additionally, we use a procedure called progressive distillation to accelerate the generation time of our method, which is typically a significant challenge for diffusion models despite their state-of-the-art precision.

High Energy Physics - Phenomenology,High Energy Physics - Experiment

What problem does this paper attempt to address?

The paper attempts to address the problem of rapidly generating point cloud data in high-energy physics, particularly the formation process of jets produced by proton-proton collisions. Traditional methods face several challenges when handling this type of data:

1. **Data Characteristics**: Particle physics datasets are typically described by continuous coordinates (unlike grid points in images) and possess symmetries such as permutation invariance, along with random dimensions. Therefore, standard deep generative models (such as those generating images or fixed feature sets) perform limitedly in this context.

2. **Computational Efficiency**: Physics-based simulations, while accurate, are computationally expensive, becoming a bottleneck in processing large-scale datasets. Classical fast approximation methods exist but often lack the expressive power to achieve the high fidelity of full simulations in certain cases.

3. **Limitations of Generative Models**: Existing generative models (such as GANs and autoencoders) are constrained by pixelation when handling jet formation, as jets are naturally represented as variable-sized point clouds, and projecting them onto a fixed-size grid loses spatial location information.

To address these issues, the paper proposes a novel neural network simulation method based on diffusion models, called Fast Point Cloud Diffusion (FPCD). This method can generate jets with complex properties and accelerates generation time through a technique called progressive distillation, significantly improving computational efficiency while maintaining high accuracy.