Optimal Observables for the chiral magnetic effect from Machine Learning
Yuji Hirono, Kazuki Ikeda, Dmitri E. Kharzeev, Ziyi Liu, Shuzhe Shi · 2025
Read Optimal Observables for the chiral magnetic effect from Machine Learning by Yuji Hirono, Kazuki Ikeda, Dmitri E. Kharzeev, Ziyi Liu, Shuzhe Shi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03248.txt.
Introduction
Yuji Hirono and four co-authors build optimal observables for the chiral magnetic effect with machine learning in a 2025 paper. The effect is a current that a strong magnetic field should drive in hot quark matter. In heavy-ion collisions, its faint signal drowns in background flows that mimic it.
The team write broad observables from flow harmonics, then let an optimizer tune them for the best signal-to-background split. On simulated collisions from a standard fluid model, the learned measures beat the classic correlators by a wide margin. Background stays low while the gain holds. The method gives experiments a sharper tool for a long-sought effect.
Catalog record
| CREATOR | Yuji Hirono, Kazuki Ikeda, Dmitri E. Kharzeev, Ziyi Liu, Shuzhe Shi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-03248
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Optimal Observables for the chiral magnetic effect from Machine Learning Yuji Hirono, Kazuki Ikeda, Dmitri E. Kharzeev, Ziyi Liu, Shuzhe Shi Abstract The detection of the chiral magnetic effect (CME) in relativistic heavy-ion collisions remains challenging due to substantial background contributions that obscure the expected signal. In this Letter, we present a novel machine learning approach for constructing optimized observables that significantly enhance CME detection capabilities. By parameterizing generic observables constructed from flow harmonics and optimizing them to maximize the signal-to-background ratio, we systematically develop CME-sensitive measures that outperform conventional methods. Using simulated data from the Anomalous Viscous Fluid Dynamics framework, our machine learning observables demonstrate up to 90% higher sensitivity to CME signals compared to traditional γ \gamma and δ \delta correlators, while maintaining minimal background contamination. The constructed observables provide physical insight into optimal CME detection strategies, and offer a promising path forward for experimental searches of CME at RHIC and the LHC. \CJKencfamily UTF8mc I Introduction The chiral magnetic effect (CME) is a macroscopic quantum phenomenon in which an electric current is generated along an external magnetic field in systems with chirality imbalance [ 1 , 2 , 3 ] . As a direct manifestation of the chiral anomaly in quantum field theory, the CME has profound implications across particle physics, nuclear physics, condensed matter physics, and cosmology. In relativistic heavy-ion collisions, where intense magnetic fields ( 10^{13}-10^{15} Tesla) are transiently created perpendicular to the reaction plane, the CME is expected to induce an electric charge separation that could provide experimental evidence for QCD topological transitions. Despite more than a decade of experimental efforts at RHIC and the LHC, unambiguous detection of the CME in heavy-ion collisions remains elusive [ 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 ] . The primary challenge arises from substantial background effects that can conspire to produce similar experimental signatures [ 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 ] . These background contributions arise from a complex interplay of local charge conservation, transverse momentum conservation, resonance decays, and collective flow phenomena, particularly elliptic flow ( v_{2} ), see e.g., [ 30 , 31 , 32 , 33 , 34 , 35 , 24 ] for reviews. Conventional observables for CME detection, such as the \gamma and \delta correlators [ 19 ] , measure charge-dependent azimuthal correlations but suffer from significant background contamination. These correlators cannot effectively disentangle CME-induced charge separation from background correlations, leading to inconclusive results despite extensive measurements. Recent experimental strategies, including the isobar collision program at RHIC, have attempted to address this challenge by comparing systems with different CME signals but similar backgrounds, yet definitive evidence remains pending [ 36 ] . Machine learning has emerged as a powerful approach for addressing complex data analysis challenges in high-energy nuclear physics. Recent studies have demonstrated that deep learning techniques can identify subtle patterns in heavy-ion collision data that may escape detection by traditional observables. The ability of machine learning algorithms to extract complex, non-linear relationships from high-dimensional data makes them particularly promising for the challenging task of CME detection [ 37 , 38 , 39 , 40 ] . In this Letter, we introduce a novel machine learning approach to CME detection that combines the power of optimization algorithms with physical understanding of the underlying phenomena. Rather than employing a black-box neural network, we parametrize a general class of observables constructed from flow harmonics and systematically optimize their coefficients to maximize sensitivity to CME signals while suppressing background contributions. This approach yields physically interpretable observables that significantly outperform traditional correlators in discriminating between signal and background. Our method builds on recent theoretical advances in CME-related observables and leverages experimental insights from current measurements. We demonstrate its effectiveness using the Anomalous Viscous Fluid Dynamics (AVFD) simulation framework [ 41 , 42 , 43 ] , which provides a realistic hydrodynamic description of heavy-ion collisions incorporating both CME signals and background effects. The resulting machine learning observables offer a promising new direction for experimental CME searches that could finally provide conclusive evidence for this fundamental quantum phenomenon. II method and results We start by defining the harmonic coefficients weighted by powers of the transverse momentum ( p_{T} ), \displaystyle v_{n}^{(m)}\equiv\frac{\langle p_{T}^{m}\,\cos(n\,\varphi_n)\rangle}{\langle p_{T}^{m}\rangle}\,,\quad a_{n}^{(m)}\equiv\frac{\langle p_{T}^{m}\,\sin(n\,\varphi_n)\rangle}{\langle p_{T}^{m}\rangle}\,, (1) with all azimuthal angles \varphi being defined with respect to the reaction plane. We then separate them into P-odd and P-even observables — the former includes all harmonic coefficients of odd order, \displaystyle\begin{split}\{o\}\equiv\left\{\left(a_{2j-1,c}^{(m)},v_{2j-1,c}^{(m)}\right)\right\}_{\begin{subarray}{c}j\in\mathbb{N},\,0<j\leq N/2,\\ m\in\{0,\ldots,M\},\\ c\in\{\pi^{+},\pi^{-}\}\end{subarray}}\;,\end{split} (2) while the latter includes the even-order ones, and in addition the inverse charge multiplicity of the event: \displaystyle\begin{split}\{e\}\equiv\left\{\frac{1}{N_{\mathrm{ch}}}\right\}\bigcup\left\{\left(a_{2j,c}^{(m)},v_{2j,c}^{(m)}\right)\right\}_{\begin{subarray}{c}j\in\mathbb{N},\,0<j\leq N/2,\\ m\in\{0,\ldots,M\},\\ c\in\{\pi^{+},\pi^{-}\}\end{subarray}}\;.\end{split} (3) Then we construct a potential CME observable (which should be P-even to survive the event averaging), that incorporates all terms that are linear and bilinear in P-even observables, as well as bilinear in P-odd ones: \displaystyle\begin{split}O=\;&\sum_{j}X_{j}^{(\mathrm{L})}\,e_{j}+\sum_{j,j^{\prime}}X^{(\mathrm{E})}_{j,j^{\prime}}\,e_{j}\,e_{j^{\prime}}+\sum_{j,j^{\prime}}X^{(\mathrm{O})}_{j,j^{\prime}}\,o_{j}\,o_{j^{\prime}}\,.\end{split} (4) It is straightforward to see that the conventional CME observables, the \gamma and \delta correlators, represent special cases of Eq. ( 4 ). We model heavy-ion collisions using the Event-by-Event (EBE) Anomalous Viscous Fluid Dynamics (AVFD) simulation framework [ 41 , 42 , 43 , 44 ] , which realistically simulates both the CME signal and the local charge conservation background (LCC)–the latter being a dominant source of contamination in conventional observables. We have generated \sim 3\times 10^{6} events for 30-40\% centrality cut of \mathrm{Au}+\mathrm{Au} collisions at \sqrt{s_{\mathrm{NN}}}=200~\mathrm{GeV} , for the four scenarios when both CME and LCC are separately turned on and off. When activated, the CME and LCC parameters are set to n_{5}/s=0.1 and P_{\mathrm{LCC}}=0.33 , respectively. The lifetime of the magnetic field is set to 0.6~\mathrm{fm} in the simulation. Given a parameter set \left\{X_{j}^{(\mathrm{L})},X^{(\mathrm{E})}_{j,j^{\prime}},X^{(\mathrm{O})}_{j,j^{\prime}}\right\} , we calculate the means and standard deviations of O for all four data sets, denoted as \displaystyle O_{\mathrm{C},\mathrm{L}}=\langle O\rangle^{\mathrm{CME-on}}_{\mathrm{LCC-on}}\,,\quad\Delta_{\mathrm{C},\mathrm{L}}=\big(\langle O^{2}\rangle^{\mathrm{CME-on}}_{\mathrm{LCC-on}}-O_{\mathrm{C},\mathrm{L}}^{2}\big)^{\frac{1}{2}}\,, (5) and likewise we define O_{\mathrm{C},{\mathrm{L}\!\!\!\!-}} , O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}} , O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}} and their corresponding standard deviations ( \Delta ’s), when one or both of CME and LCC are turned off. Our goal is to construct an observable that maximizes sensitivity to the CME signal while minimizing responses to non-CME backgrounds, thereby addressing the primary challenge in current experimental efforts. To achieve this, we explicitly maximize the statistical significance of CME-induced differences while simultaneously minimizing the significance of background-induced variations, using the following loss function \displaystyle\begin{split}\mathcal{L}=\;&\lambda\;\bigg(\frac{(O_{{\mathrm{C}},{\mathrm{L}}}-O_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}})^{2}}{\Delta^{2}_{{\mathrm{C}},{\mathrm{L}}}+\Delta^{2}_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}}}+\frac{(O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}}-O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}})^{2}}{\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}}+\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}}}\bigg)\\ &-\bigg(\frac{(O_{{\mathrm{C},{\mathrm{L}}}}-O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}})^{2}}{\Delta^{2}_{{\mathrm{C},{\mathrm{L}}}}+\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}}}+\frac{(O_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}}-O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}})^{2}}{\Delta^{2}_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}}+\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}}}\bigg)\,,\end{split} (6) where the relative factor \lambda is taken as 10^{4} . We then iterate the parameters \left\{X_{j}^{(\mathrm{L})},X^{(\mathrm{E})}_{j,j^{\prime}},X^{(\mathrm{O})}_{j,j^{\prime}}\right\} using the gradient descent method and minimize \mathcal{L} . Then we compute projection of significance with N_{\mathrm{ev}}=3\times 10^{6} events by \displaystyle S=\frac{f_{c}}{2}\left(\frac{O_{{\mathrm{C},{\mathrm{L}}}}-O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}}}{\sqrt{\frac{\Delta^{2}_{{\mathrm{C},{\mathrm{L}}}}+\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}}}}{N_{\mathrm{ev}}}}}+\frac{O_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}}-O_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}}}{\sqrt{\frac{\Delta^{2}_{{\mathrm{C}},{\mathrm{L}\!\!\!\!-}}+\Delta^{2}_{{\mathrm{C}\!\!\!\!-},{\mathrm{L}\!\!\!\!-}}}{N_{\mathrm{ev}}}}}\right)\,. (7) We have included an extra correction factor of f_{c}=\frac{1}{4} in the estimation of signal significance to tune the assumed value of n_{5}/s 1 1 1 While we start from a simulation date set with axial-charge-to-entropy ratio n_{5}/s=0.1 , it is found that this ratio should be chosen as 0.05 to fit the experimental data that f_{\mathrm{CME}}\equiv\gamma^{\mathrm{OS-SS}}_{\mathrm{CME}}/\gamma^{\mathrm{OS-SS}}\approx 0.15 [ 11 ] . The signal significance should be corrected by the factor f_{c}=(0.05/0.1)^{2} . . Upon a linear response analysis of both CME