FCC Tau Reconstruction - November 20, 2026
Present: Laurits Tani, Michel Hernandez Villanueva, Mohadeseh Dadashzadeh, Olmo Arquero, Karl Ehataht, Norman Seeba
CMS Hadron-Plus-Strips (HPS) algorithm and tuning for FCC (Laurits)
- Overview of the CMS HPS tau reconstruction algorithm and its evolution from Run 1 to Run 2+
- Charged hadrons and electrons form prongs; photons and electrons are clustered into 𝜂-𝜙 “strips” to reconstruct 𝜋0 decays (wide in 𝜙 due to bending in the strong magnetic field)
- Constructs decay mode hypotheses (1-prong, 1-prong+1𝜋0, 1-prong+2𝜋0, 3-prong, 3-prong+1𝜋0) targeting intermediate resonances (𝜌, 𝑎1)
- Applies kinematic, mass-window, and signal-cone cuts; surviving candidates are ranked by charge multiplicity (3-prong over 1-prong), 𝑝𝑇, number of strips, and isolation
- HPS handles candidate reconstruction (decay mode and visible 4-momentum), whereas tau identification was later decoupled and delegated to ML classifiers (DeepTau)
- Key design updates across CMS versions
- Transitioned from fixed to 𝑝𝑇-dependent dynamic strip sizes to account for photon conversions and collimation at high 𝑝𝑇 in busy 𝑝𝑝 environments
- Tight cuts initially reduced fake rates at the cost of ~50% efficiency; cuts were later loosened once DeepTau was introduced on top
- Re-implementation and hyperparameter optimization for FCC-ee
- Re-implemented CMSSW HPS in Python using Numba for fast execution
- Evaluated on 270,000 jets from 𝑍→𝜏𝜏 full simulation samples with the CLD detector geometry (half for optimization, half for evaluation)
- Tuned 50 of the 84 total HPS parameters (23 decay mode cuts, 18 strip building, 6 signal cone, 3 global) using Optuna (Tree-structured Parzen Estimator) across 5,000 trials and 8 random seeds
- Two concurrent optimization objectives: decay mode classification accuracy (across 5 modes) and visible momentum resolution
- Optimization results and performance
- Significant improvement in decay mode accuracy and a twofold improvement in momentum resolution compared to default CMSSW parameters, approaching the ideal MC-truth matching limit
- Notably recovered classification performance for the 1-prong+2𝜋0 decay mode (previously misclassified due to strip merging in default settings)
- fANOVA parameter importance analysis identified minimum strip 𝑝𝑇 among the most impactful parameters
Questions/comments
- How does the retuned HPS algorithm perform on background rejection (e.g., 𝑍→𝑞𝑞¯), given that only 𝑍→𝜏𝜏 samples were used in the tuning?
- Not yet tested, as looser cuts are viable at FCC-ee where ML taggers handle background separation well, though Laurits will run the retuned HPS on a 𝑍→𝑞𝑞¯ sample to check candidate survival rates
- Comparison with existing FCC heuristic approaches (Sofia and Olmo’s cone-based algorithms)
- Those approaches resemble early CMS cone-based tau reconstruction; in an informal test, Olmo’s algorithm outperformed retuned HPS in efficiency due to HPS’s tighter cuts, though a fair comparison requires checking background rates
- Is there any constraint placed on the median of 𝑝reco/𝑝vis (or 𝑝gen) to stay close to 1 and avoid momentum scale biases?
- No explicit constraint was enforced on the median during optimization
- Follow-up / Next steps
- Michel to review CMS detector features in relation to the HPS design choices to understand which aspects translate to FCC detector concepts
- Laurits to evaluate 𝑍→𝑞𝑞¯ background survival rates with the tuned parameters
AOB
- Open call for volunteers to present overviews of other existing reconstruction algorithms in future meetings
There are minutes attached to this event.
Show them.