Speaker
Description
In relativistic heavy-ion collisions, the collision centrality is conventionally determined by sorting events into percentile bins of charged-particle multiplicity.
The procedure is applied to a stochastic relation between multiplicity and initial geometry.
The associated finite volume fluctuation is therefore irreducible and introduce an inevitable uncertainty, impacting many centrality-dependent measurements.
We present a machine learning approach to centrality determination: the Inference-driven Participant Determination (IPD) method.
It replaces the hard multiplicity cut with Bayesian probabilistic inference.
The method uses the Monte-Carlo Glauber (MCG) model as a prior and a product of negative binomial distributions as the likelihood, yielding an event-by-event posterior distribution of participant numbers ($N_{\rm part}$) and binary collisions ($N_{\rm coll}$) directly from final-state charged-particle multiplicities.
Unlike supervised learning approaches, IPD requires no transport-model training and is directly applicable to experimental data.
A self-consistent closure test using a UrQMD-MCG hybrid model at $\sqrt{s_{\rm NN}} = 19.6$~GeV demonstrates that the method produces well-calibrated posteriors with
negligible bias.
The IPD-based centrality yields $\sigma(N_{\rm part})$ distributions that are 11\% to 46\% narrower than those from conventional multiplicity-based classification.
This method suppresses volume fluctuation effects, allowing cleaner measurements with less affected by contamination from volume fluctuations.
The improvement in centrality purity is robust across $\sqrt{s_{\rm NN}} = 3.0$, 7.7, and 19.6~GeV, covering the RHIC Beam Energy Scan range.
The IPD framework is general and can be adapted to centrality determination in other heavy-ion experiments, such as CBM and CEE.
