Classification of flavor dependence of Chiral Magnetic Effect with Deep Neural Network using multiple correlators

Abstract

We study the flavor dependence of the Chiral Magnetic Effect (CME) by analyzing two key charge-separation correlators used to characterize the charge separation effect: the conventional $\Delta\gamma$ and the recently proposed $R_{\psi_2}$. Using the AMPT (A Multiphase Transport) model with an initial-state centrality-dependent charge separation, we evaluate the sensitivity of these correlators to 2-flavor ($u,d$) and 3-flavor ($u,d,s$) quark scenarios. While both correlators exhibit modest flavor dependence in mid-central (30-50\%) collisions, their discriminative power varies significantly with centrality and transverse momentum ($p_T$), limiting their utility disentangling the flavor dependent scenarios. To overcome these limitations, we develop a neural network classifier trained on final-state hadronic observables (e.g., $dN_{ch}/d\eta$, $p_T$ spectra). The model achieves $>90\%$ accuracy in flavor classification by leveraging multi-observable correlations, with $p_T$-differential features proving particularly discriminative. Crucially, by incorporating background contributions directly into the training data, our approach provides more reliable flavor estimates than correlator-only methods.