High-precision regressors for particle physics.

Area of Science:

• High-energy physics
• Computational physics
• Machine learning

Background:

• Monte Carlo simulations are crucial for particle collider physics but computationally intensive.
• Current simulation methods require significant computational resources, limiting research speed.

Purpose of the Study:

• To develop high-precision machine learning regressors for physics simulations.
• To reduce the computational burden of Monte Carlo simulations at particle colliders.
• To improve the efficiency of data generation for collider experiments.

Main Methods:

• Tuning various machine learning regressors for high-precision requirements (<1% relative error).
• Leveraging particle physics symmetry arguments to optimize regressor performance.
• Designing a Deep Neural Network with skip connections, inspired by ResNets.

Main Results:

• Achieved significant speedups (100x-1000x) compared to first-principles computations.
• Reduced the number of required regressors by an order of magnitude using symmetry arguments.
• Demonstrated that boosted decision trees outperform neural networks in lower dimensions, while neural networks excel in higher dimensions.

Conclusions:

• Machine learning regressors can substantially decrease the computational cost of Monte Carlo simulations.
• The developed methods offer a path to significantly reduce training and storage burdens for current and future collider experiments.
• Optimized machine learning models, incorporating physics knowledge, are key to advancing particle physics research.