Machine Learned Particle Detector Simulations
The use of machine learning algorithms is an attractive way to produce very fast detector simulations for scattering reactions that can otherwise be computationally expensive. Here we develop a factorised approach where we deal with each particle produced in a reaction individually: first determine if it was detected (acceptance) and second determine its reconstructed variables such as four momentum (reconstruction). For the acceptance we propose using a probability classification density ratio technique to determine the probability the particle was detected as a function of many variables. Neural Network and Boosted Decision Tree classifiers were tested for this purpose and we found using a combination of both, through a reweighting stage, provided the most reliable results. For reconstruction a simple method of synthetic data generation, based on nearest neighbour or decision trees was developed. Using a toy parameterised detector we demonstrate that such a method can reliably and accurately reproduce kinematic distributions from a physics reaction. The relatively simple algorithms allow for small training overheads whilst producing reliable results. Possible applications for such fast simulated data include Toy-MC studies of parameter extraction, preprocessing expensive simulations or generating templates for background distributions shapes.