Energy spectra of elemental groups of cosmic rays with the KASCADE experiment data and machine learning

We report the reconstruction of the mass component spectra of cosmic rays (protons, helium, carbon, silicon and iron) and their mean mass composition, at energies from 1.4 to 100 PeV. The results are derived from the archival data of the extensive air shower experiment KASCADE. We use a novel machine learning technique developed specifically for this reconstruction, and post-LHC hadronic interaction models: QGSJet-II.04, EPOS-LHC and Sibyll 2.3c. We have found an excess of the proton component and a deficit of intermediate and heavy nuclei components compared to the original KASCADE results. The spectra of protons and helium show a knee-like behavior at ~ 4.4 PeV and ~ 11 PeV, with significances 5.2${\sigma}$ and 3.9${\sigma}$, respectively. The spectrum of the iron component has a hint (2.4${\sigma}$) of a hardening at ~ 4.5 PeV, which can be interpreted as a counterpart of a hardening in the proton spectrum at 166 TeV, recently reported by the GRAPES-3 experiment. The systematic uncertainties of our analysis were found to be smaller than those of the original KASCADE, as well as those of IceTop and TALE experiments, over the most part of the energy range studied. We also estimated separately the uncertainty related to the difference between the three mentioned hadronic interaction models. We also compute a mean logarithm mass of cosmic ray flux as a function of energy. It is in agreement with the results of IceTop, TALE and LHAASO within the uncertainties.