Search for a command to run...
Generative Adversarial Networks have been used extensively to generate high resolution images in different fields. The present work explores the adversarial paradigm to train a deep neural network as a tool for scientific simulation, more specifically, simulation of High Energy Physics detectors. At the core of this approach lies the possibility to interpret the detector response as an image: the main challenges being represented by the large pixel intensity dynamic range (spanning over more than ten orders of magnitude) together with a high level of image sparsity. An initial prototype [1] simulated electrons travelling through an example electromagnetic calorimeter and depositing energy in its volume. It achieved very good results using an adversarial approach to reproduce a simplified physics use case. The present work extends the model to a more complex and realistic scenario: the conditioning approach is modified to include additional physics variables, the network architecture is adjusted to take into account larger image size and more complex pixel dynamic features, the cost function is re-designed in order to include physics based constraints. We ntroduce a multi-step training process based on transfer learning, by initially optimizing the network performance on a restricted particle energy range (therefore, a simplified pixel intensity distribution). Further training the network on the full energy range yields very good results, without the need of additional optimization. Beyond a qualitative visual inspection, the validation of GAN performance is based on a detailed comparison to standard Monte Carlo simulation in terms of several physics quantities: results show a remarkable agreement, ranging from a few percent up to 10% across a large particle energy range.