Faculty Sponsor: Dr. Michelle Kuchera
Machine Learning for Supersymmetry Searches Braden Kronheim, Alexander Karbo, Dr. Michelle Kuchera, Dr. Raghu Ramanujan
The current fundamental theory for particle physics is the Standard Model. While the Standard Model has been a very successful theory, it is not without its problems and physicists are constantly looking for ways to extend it. One such extension is supersymmetry, which predicts the existence of a partner particle for each of known fundamental particles which should be observable in particle accelerators. Thus, calculating the probability of creating these particles in accelerators is of supreme importance for verifying the theory. Unfortunately, the full theory has 105 free parameters and existing prediction programs can only generate about 16.6 points per cpu hour. In this project the use of machine learning was studied to make predictions from a simplified theory, the phenomenological Minimal Supersymmetric Standard Model, which only has 19 free parameters. Theoretically, a dense neural network should be able to simulate the output of an arbitrarily complex function, making neural networks optimal for this task. Within the project standard networks which have definite values were compared to networks with distributions for values and to chains of networks obtained through a Markov Chain. The networks were compared through their ability to accurately predict the probabilities and have a wide variance in output when unsure. Unsurprisingly, the networks with distributions for values outperformed the networks with definite values, and the networks formed from Markov Chains were the best. The best collection of networks created yet was able to reproduce the particle probability with a percent error of 8.1% and with 94.08% of real points within 3 standard deviations of the output distributions.