High-precision measurements of proton and helium nuclei obtained from direct cosmic ray experiments provide valuable insight into the mechanisms of CR acceleration and propagation in the Galaxy.
The space-based calorimeters, CALET as well as DAMPE, have recently revealed an additional spectral feature at tens of TeV, i.e a softening of the flux, which is not predicted by traditional cosmic-ray models.
However, the energy range in the multi-TeV region is still subject to large uncertainties, mainly due to the limited statistics
of the available data. In this work, we present an analysis based on Boosted Decision Trees (BDT) to improve the statistical precision
of light element fluxes by extending the fiducial acceptance of CALET data. The performance of BDT-based selections is evaluated and compared with standard and other machine-learning techniques, focusing mainly on helium nuclei.