Geant4 open points

[freddieknets]

TODO (short-term):

• Tracking:
  • Check the preselection in C (and make it optional)
  • Double-check post-processing (virtual energy can result in negative numbers!)
  • Need a direct benchmark to a pure Geant4 model and tracking
  • Ions have issues: very weird return distribution (see dedicated issue)
  • Returned masses have weird variations (for the same particle type). This should not happen. Investigate and fix
  • test_geant4_many_particles.py::test_lead[hit-geant4] fails sometimes for (acc ion loss not accounted for, or something like that?)
  • Jaw accuracy is up to 5e-9 (bad), though only a few random particles (and only on the inside) following a tilted slope in (x, px)
  • Should _accumulated_ion_loss not be a collimator property, but use a dedicated particle state (one particle per collimator representing this)?
  • Upstream s position is shifted 250um. Is this valid? Double-check, and verify downstream s position
• Implement Geant4Crystal
• Implement Geant4Block
• Should Geant4CollimatorTip really be a separate element?
• Make compatible with prebuild kernels: how to resolve the dependency on pybind11?
• Implement BdsimCollimator in pure python using the new python links in BDSIM
• Code compatibility with new Material
• Return particle type selections
• Better control over physics processes in Geant4, return types, and cuts etc
• Two engines in parallel (e.g. some FLUKA elements and some Geant4 elements) cause issues (different cwd)!
• Sometimes many_particles test fails as the energy does not sum while there is no Eacc, Evirtual, Emassless, nor Eout. This is weird and not understood. Example (100 Pb208):

 >           assert np.isclose(E0*alive_before, Edead + Eacc + Evirtual + Emassless + Eout, atol=100*tol, rtol=tol)
E           assert np.False_
E            +  where np.False_ = <function isclose at 0x1118bd430>((np.float64(557600033653993.8) * np.int64(100)), ((((LinkedArrayCpu(5.55454199e+16) + np.float64(0.0)) + LinkedArrayCpu(0.)) + LinkedArrayCpu(0.)) + LinkedArrayCpu(0.)), atol=(100 * 1e-12), rtol=1e-12)
E            +    where <function isclose at 0x1118bd430> = np.isclose

TODO (long-term):

• Translate C++ code (previous collimasim) into extern C { } to avoid dependency on pybind11 and get smooth Xsuite-C
• Implement generic geometry in GMAD

Ideas:

• Can we make the Geant4 elements OpenMP-compatible?
• Maybe the engines should not be global, such that multiple connections can be ran in parallel in the same Python session? As a solution for OpenMP?