CVUT clusters

Research infrastructure Fermilab-CZ contributes to operation of two clusters for statistical calculations and machine learning, numerical server HELIOS and testing statistical server VKSTAT located at Department of Mathematics of Faculty of Nuclear Sciences and Physical Engineering of Czech Technical University (FJFI CVUT) in Prague.


Statistical server VKSTAT

  • 64 cores and 350 GPUs
Independent server VKSTAT is mainly used for processing and classification of Monte-Carlo HEP simulations, running computationally demanding statistical algorithms based on robust methods of stochastic models parameters estimation, identification of HEP models and their quality testing including homogeneity tests of Monte-Carlo vs. real data from experiments (NOvA)DUNE/Fermilab, ATLAS/CERN, protoDUNE/CERN, STAR/BNL.

It also uses implemented advanced algorithms for non-parametric probability density estimates, e.g. estimates of distribution mixtures and transformed multidimensional core estimates based on minimization of divergent D-distance.

Numerical server HELIOS

  • 672 cores, more details here
Numerical server HELIOS is primarily used for demanding tasks of optimization of extensive structured convolutional neural networks for signal binary classification. It is also used for processing of large data files from neutrino oscillations, proton-proton collisions, heavy elements nuclei collisions with parallel machine algorithms (NN, DNN, CNN, CVN, ResNET, TransferLearning ...) executed on powerful graphic cards. The algorithms are used by Fermilab experiment NOvA/DUNE for 3D reconstruction and separation of detected neutrino tracks in both near and far detector and STAR/BNL experiment for processing of d+Au and Au+Au decays. Machine calculations can be done in various environments, e.g. Matlab, R, Python, MPI/hybrid OpenMPI or CUDA jobs. Both clusters allows for parallel calculations for high-dimentional problems of particle physics and have highspeed connections with FZU farm and Fermilab.

Both servers are running SW tools for use in decision and classification trees (e.g. divergent supervised decision trees) and statistical algorithms for reduction of physics analysis dimensionality and homogeneity testing of physics variables for Monte-Carlo HEP files.

Both servers are in open access mode for various users and collaborating phd and diploma students from various field, e.g. transportation, material defectoscopy, sociological research etc.




Architecture of convolutional neural network SE-ResNet34 for image data used for classification of neutrino flavors and other properties.


Helios specs