Learning to simulate high energy particle collisions from unlabeled data.

Area of Science:

• Particle Physics
• Statistical Inference
• Computational Science

Background:

• Simulations are crucial for mapping theoretical models to experimental data in scientific fields.
• Analytical descriptions of experimental data reconstruction from indirect measurements are often challenging.
• Current numerical simulations are computationally expensive.

Purpose of the Study:

• Introduce Optimal-Transport-based Unfolding and Simulation (OTUS), a novel, fast simulator.
• Enable direct transformation from theoretical models to experimental data using unsupervised machine learning.
• Provide a potential replacement for computationally costly simulation methods.

Main Methods:

• OTUS utilizes unsupervised machine learning, specifically a probabilistic autoencoder.
• The autoencoder is trained to directly transform between theoretical models and experimental data.
• The latent space of the autoencoder is identified with the space of theoretical models.

Main Results:

• OTUS demonstrates capability in predicting experimental data from theoretical models.
• Proof-of-principle results are shown for Z-boson and top-quark decays in particle physics.
• The decoder network acts as a fast, predictive simulator.

Conclusions:

• OTUS offers a computationally efficient alternative to traditional simulation methods.
• The framework has broad applicability across various scientific disciplines relying on statistical inference.
• This approach accelerates the process of testing theoretical models against experimental data.