Data Analysis and Machine Learning at the Computational Science Group

In this talk I will summarize the current activities of the computational science group at HZDR, which range from establishing research data publication (RODARE) and management platforms to providing numerical and computational expertise in various research projects. The latter aspect will be presented in detail through two selected ongoing projects.

In a project with the institute for resource ecology at HZDR, we developed a framework for the analysis of spectra of mixed solutions. The goal of the analysis is to discover how many species are in a given sample and in what concentration while at the same time extracting their unknown spectrum. A number of numerical techniques can be employed to this end, each requiring different amounts prior knowledge and different types of measurements. Here the primary task of the analysis framework is to unify a zoo of different implementations of similar methods and making all methods available to all scientists. Additionally, it enables simple use of remote computing resources, which allow for more computationally intensive analysis which can add a more reliable way to estimate confidence bounds.

In another project, we are using deep learning approaches to develop an automated safety system for the high-power laser systems DRACO and PENELOPE at HZDR. Here the goal is to detect defects or scatterers which focus parts of a yet unfocussed beam. These can, when left unchecked, cause cascades of failing mirrors, lenses, and non-linear crystals and should thus be detected in the time between two shots. This work uses deep convolutional neural networks implemented through the Caffe framework to achieve real-time detection and localization of impurities in the beam profile.