Speaker
Description
The chiral magnetic effect (CME) in relativistic heavy-ion collisions induces a charge separation along the magnetic field, yet its detection is obscured by large backgrounds, especially from local charge conservation (LCC). Using event-by-event AVFD simulations of Au+Au collisions at $\sqrt{s_{NN}}=200$ GeV, we studied machine-learning strategies to enhance CME discrimination.
First, we construct observables as linear and bilinear combinations of harmonic coefficients and optimize their coefficients via gradient descent. The resulting observables suppress backgrounds to near-zero levels and improve CME sensitivity by up to 110% relative to conventional $\gamma$ correlators.
Second, by exploiting a multilayer perceptron neural network, we construct “optimal CME observables” as non-linear functions of C-even, P-even harmonic coefficients. We observe that linear combinations of P?even observables dominate the discrimination, as the network's performance (achieving ~74% classification accuracy) closely matches the theoretical sensitivity derived from linear and bilinear combinations of such observables, indicating that linear contributions are the main source of separability while nonlinear gains are marginal.
Our framework offers a systematic, interpretable route to CME detection, and the optimized observables are directly applicable to experimental data at RHIC and LHC, paving the way toward finding CME .
