With the steady increase in accuracy, size and complexity of astroparticle physics experiments the need for an extensive amount of high precision Monte Carlo simulations is rapidly growing. Contrary to the increasing demand is the demise of ``Moore's law'' which leads to situations where the system structure of high-performance computing is fundamentally changing and large amounts of money are invested in new infrastructure. In the field of astroparticle physics CORSIKA 7 is currently the most commonly used simulation program, therefore the presented work is focused on this software. All methods provided are also currently transferred to the new CORSIKA 8 framework which will replace CORSIKA 7 in the near future. Due to various constraints on hardware, geometry and physics no experiment is able to observe the full air shower particle cascade developing in the atmosphere. The removal of the non-visible phase space of the cascade at an early stage of the simulation has immense potential to reduce the expense of calculations without changing the results of the simulation for the experiment. Fast machine learning models allow the identification and removal of particles from those regions to speed up the simulations. This works for example for neutrinos by orders of magnitude. First results are shown to demonstrate this technique.
Furthermore, when showers are simulated with the IACT configuration around 75\% of the time is spent on the Cherenkov photon creation and propagation. We also show results from parallelizing this part of the simulation on GPUs and CPUs with OpenCL.