Revised TOPO model for hh->bbWW->bbmunuqq

[LukasEbeling]

This PR adds a revised model (versions 2 and 3) including the following changes:

• hwQual of muons added as features
• etaAtVtx and phiAtVtx are used instead of eta and phi for muons
• phi coordinate of HT removed
• better balance between signal and background events during training
• model trained with 10% dropout to avoid overtraining

For model v3 the architecture was revised to meet the stringent latency requirements. Instead of three hidden layers, the model included two layers with 128 + 64 nodes.

The model was trained on 2025 MC samples:

• Single Neutrino Events (background)
• Di-Higgs events, single muon final state (signal)

[LukasEbeling]

High efficiency for loose, medium and tight working point (defined below). Comparable or better efficiency than best five cut-based triggers combined.

Working points defined based on pure rate (zero bias events not triggered by any other unprescaled trigger):

working point	score	pure rate [kHz]	total rate [kHz]	efficiency [%]
loose	>= 1016	1.0	7.6	97
medium	>= 1018	0.5	5.4	96
tight	>= 1021	0.1	2.5	93

The rate are calculated using a zero bias dataset. The efficiencies are determined based on a test dataset (not included during the model training).

[missirol]

I don't think it's okay to use hard-coded paths to /cvmfs (otherwise, how does one build this package on nodes without /cvmfs access, and/or using the latest version of the hls or hls4mlEmulatorExtras externals ?).

For comparison, see here what is done for AXOL1TL.

If so, this should be fixed everywhere in this package (and the fix should be propagated to all the relevant cmsdist branches, i.e. 16_0_X and 16_1_X as of now).

[artlbv]

hi Marino, thank you for looking into this!
I made the changes here: LukasEbeling#2

[missirol]

Thanks, @artlbv.

[LukasEbeling]

Good agreement between the keras model and the cmssw output score: