R&D task number: G4RD7
Development and integration of deep Learning Inference for Fast Simulation Applications
Within the simulation chain, the particle transport in calorimeter sub-detectors represents the biggest fraction of the CPU usage during the simulation run. This is due to a large number of secondaries produced within the volumes and often exacerbated by complicated geometries (especially for the high-granularity calorimeters). On the other hand, the exact tracking of all the particles present in the volumes is not relevant and only the measure of the combined energy deposition in the calorimeter cells is needed. Thus, the output of the calorimeter shower simulation can be interpreted as an image, where calorimeter cells correspond to pixels. This analogy suggest that one can use generative Machine Learning models as a fast simulation technique, where detailed tracking of a particle inside a sub-detector is replaced by a trained network capable of producing the detector response to the given incoming particle.
The challenge in data generation comes from building models able to understand and estimate a dataset's joint density p(x). Moreover, for fast simulation generative models additional challenges arise from the large array of possible data configurations, the need for flexible models, able to represent complex densities, ability to capture nonlinear, long-range correlations and input varying dependencies.
In order to address these challenges and construct a compelling target aware generation distribution, in the ongoing work, the simulation event is represented as volumetric data from which we learn the true underlying probability distributions of energy deposits through autoregressive models which decompose the joint density as a product of conditionals to be modelled.
To address the increase in computational costs and speed requirements for simulation related to the higher luminosity and energy of future accelerators, a number of Fast Simulation tools based on Deep Learning (DL) procedures have been developed. In order to integrate DL simulation methods with an existing Full Simulations toolkit (Geant4), we provide a light-weight, dependency-free platform for Fast Simulation Deep Learning Inference a C++ module for TensorFlow models.
The overall goal is to facilitate the usage of generative DNNs by integrating the inference module with Geant4. Thus, from a DL developer perspesctive, a trained model can easily be added to the available library of models for inference. Moreover, from a user perspective, the option for deep learning simulation would now be easily available within Geant4.
Code repository: https://github.com/ioanaif/dl-inference-module
Lead and main developers: Ioana Ifrim (firstname.lastname@example.org)