Replaces the detailed transport through the ALICE front absorber by a model, and compares the result against a full simulation of the same events.

run.sh runs both: a reference o2-sim with PIPE and ABSO only, and the same setup with the fast simulation switched on. Both use the same seed, so the two simulations see the same primaries and the comparison is paired.

The feature does nothing unless a model is named:

--configKeyValues "G4.fastSimModels=toyAbsorber;G4.fastSimEnvelope=AFaM"

G4.fastSimEnvelope is the volume the model stands in for — here AFaM, the mother of the whole absorber. G4.fastSimMinEnergy (GeV, default 1) is the threshold below which the detailed transport still runs.

toyAbsorber is a placeholder: it returns the incident particle continuing in its direction with the energy attenuated exponentially over its path through the envelope. A real absorber turns one incident hadron into a shower, so these are not physics numbers.

A model implements one function, sample(), which maps the particle that entered the region to the particles that leave it (Detectors/FastSim/include/FastSim/FastSimModel.h). The surrounding work — measuring the distance to the envelope surface, killing the incident particle, stacking what comes back, booking the energy difference as a deposit — is shared.

Geant4 consults a fast simulation model through regions, and in O2 a region can only ever be "every volume of a given material": the VMC special cuts make every logical volume a root of its own material's region, and Geant4 stops propagating a region at any such daughter. A region named after AFaM would therefore contain AFaM alone, which tracks skip entirely because its daughters touch its surface.

So the two jobs are separated. The regions are derived by walking the envelope's subtree and collecting its media, which for AFaM finds fourteen:

fast simulation: model toyAbsorber covers 14 media found under 'AFaM'
Adding fast simulation model toyAbsorber to regions ABSO_AIR0$ ABSO_AIR_ENVELOPE0$
  ABSO_CONCRETE2$ ABSO_POLYETHYLEN2$ ABSO_CARBON0$ ABSO_CARBON2$ ABSO_MAGNESIUM$
  ABSO_Ni-W-Cu0$ ABSO_Ni-W-Cu2$ ABSO_LEAD0$ ABSO_LEAD2$ ABSO_STAINLESS STEEL0$
  ABSO_STAINLESS STEEL2$

What the model measures against is the envelope itself, read from the track's own touchable. ModelTrigger also requires geometric containment in it, which is what excludes the steel support cradle — it shares its material with the end plate, so no selection by material could separate them.

Five pp minimum-bias events, Pythia8 and Geant4, -m PIPE ABSO, same seed, one EPN node against O2PDPSuite/daily-20260819-0000-1:

A 20 GeV muon fired into the muon-arm acceptance with /tracking/verbose 1 shows what the model does to a single track:

   10     54.4   -0.137     -900  1.99e+04   0.0749      287       902   AFaMgRing Transportation
   11     54.4   -0.137     -900         0 1.99e+04  4.1e+03     5e+03   AFaMgRing G4FastSimulationManagerProcess

One step of 4.1 m across the whole absorber, in place of roughly twenty muIoni steps through graphite, concrete and steel.

Not a performance result. A quarter fewer tracks buys only five percent of wall clock, because the tracks the absorber's shower contributes are cheap low-energy ones and most of the CPU in this setup is spent elsewhere. And with minimum-bias pp only the forward cone reaches the absorber at all. A CPU number worth quoting needs a workload where the absorber is on the critical path — a forward-biased generator, or the full detector where the muon arm is the point.

Not a physics validation. The toy returns one particle where a real absorber returns a shower, so the track counts above say more about the placeholder than about the absorber. Comparing the outgoing multiplicity and spectrum against a full simulation is what a trained model has to pass, and that is the measurement this example is scaffolding for.

Two further caveats worth knowing when reading any output of this: a fast step does not call the sensitive detector, so its steps disappear from MCStepLogger (harmless for a passive envelope, which has no hits); and secondaries the model creates carry TMCProcess kPNull, the code geant4_vmc gives to everything it has no VMC equivalent for, so they cannot be told apart from other tracks by process alone.