Event generation and statistical sampling for physics with deep generative models and a density information buffer.

Area of Science:

• Particle Physics
• Computational Physics
• Machine Learning

Background:

• Simulating particle physics processes demands significant computational resources, often exceeding feasible timeframes.
• Accurate simulation requires not only generating physical events but also reproducing their correct frequencies.

Purpose of the Study:

• To explore deep generative models for simulating particle physics events and their frequencies.
• To develop a faster and efficient alternative to conventional Monte Carlo generators.

Main Methods:

• Investigated deep generative models, including Variational Autoencoders, for event generation.
• Trained models on physical processes like two-body decays, e+e- → Z → l+l-, and top quark decays.
• Utilized density information from encoded Monte Carlo events to construct a sampling prior.

Main Results:

• Generative models successfully learned event generation and frequency distributions.
• Generated events showed excellent agreement with real Monte Carlo data.
• Achieved event generation several orders of magnitude faster than traditional methods.

Conclusions:

• Deep generative models offer a viable and highly efficient solution for particle physics simulations.
• This approach significantly reduces computational time while maintaining accuracy.
• Potential applications include density estimation, targeted event generation, and improved importance sampling.