Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning

arXiv:2603.29269 · hep-ph, nucl-th · Submitted 2026-03-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: We're looking at the paper: "Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning".

Mira: Here's the gist of the paper. Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning investigates the phase structure of magnetized QCD matter by determining field-dependent parameters of the Nambu-Jona-Lasinio (NJL) model through a physics-informed machine learning framework The magnetic field smoothly suppresses both G and v2, providing new microscopic insights into the response of the QCD vacuum to strong magnetic fields Theoretical Framework The Lagrangian density of the two-flavor Nambu–Jona-Lasinio (NJL) model in the presence of a background magnetic field, including the contribution from the anomalous magnetic moment (AMM) of quarks, is given by L = ψ¯ iγµDµ − mˆ + 1/2 qfκFµνσ µν ψ + G (ψψ¯)2 + (ψiγ ¯ 5 τψ)2 To determine the dynamical quark mass, the mean-field approximation is employed, leading to the constituent quark mass matrix Mˆ = ˆm + σ, where σ = −2G⟨ψψ¯ ⟩ The presence of the AMM modifies the Dirac equation for quarks to iγµDµ − Mˆ + 1/2 κqfσ µνFµν ψ = 0 The thermodynamic potential Ω is obtained following standard procedures, taking the form of a sum over Landau levels How the training signal is built The core idea involves employing machine learning as a tool for inverse engineering, using lattice QCD computed magnetic-field-dependent quark condensate data as the “ground truth” and a parameterized (field-dependent) NJL model as the “interpretable function” The neural network is trained to directly “learn” the optimal functional forms for the evolving coupling constant G(eB) and AMM ratio v2(eB) which is defined by κ(eB) = v2(eB)σ2 This involves constructing a differentiable physics-constrained module where the gap equation acts as a “hard constraint” to ensure strict adherence to fundamental physical laws Where the gains come from The optimization objective is defined by the total loss function Ltotal(θ) = λbLb + λMLM + λσLσ + λQC (LlowTQC + LhighTQC) This loss function balances data fidelity, empirical constraints, and physical priors The constituent quark mass loss LM penalizes predictions falling outside a ±5% tolerance zone of the lattice data Mlatt The physics-informed regularizer includes losses such as the Chiral Dynamical Quark Mass Monotonicity Loss Lσ, which encodes the physical behavior that σ decreases monotonically with increasing magnetic field at zero temperatures Numerical Results Within the investigated range of 0 < eB < 0.6 GeV2, both G(eB) and v2(eB) are observed to be smoothly suppressed by the external magnetic field Specifically, the coupling constant G(eB) decreases from a vacuum value of 4.805 ± 0.012 GeV−2 to 3.550 ± 0.031 GeV−2 at the highest field strength Similarly, the AMM ratio v2 is reduced from 0.771 ± 0.057 GeV−3 to 0.734 ± 0.059 GeV−3 The chiral condensate displays a complex, temperature-dependent dual behavior, capturing the inverse magnetic catalysis effect observed in lattice QCD The differentiable inversion framework establishes a general paradigm for bridging the gap between first-principles calculations and phenomenological models Future work will focus on extending the pipeline to more sophisticated effective theories, such as the Polyakov-loop NJL (PNJL) model or (2+1)-flavor systems The methodology demonstrates that machine learning is a powerful engine for discovery and parameter optimization within the context of fundamental physical laws Acknowledgments The authors thank Zhongmin Niu for insightful discussions and Minghua Wei for early contributions to this work The authors acknowledge support from the National Natural Science Foundation of China (Grant No.

Kai: 12547138) and the Anhui University of Science and Technology (Grant No.

Mira: 2025yjrc0143) X.W.

Kai: was supported by the Anhui University of Science and Technology under Grant No.

Paper summary: Kai: So we're looking at this paper titled "Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning". It sounds like they are using machine learning to figure out how things behave in strongly magnetized matter inside Quantum Chromodynamics.

Mira: Yeah, the core idea is that they're taking lattice QCD data as their ground truth—that's the actual physics result—and then using a parameterized Nambu-Jona-Lasinio model to see if it can reproduce that data.

Kai: Exactly, so the thesis is about determining two specific things: how fast the coupling constant G changes with the magnetic field, and how much that anomalous magnetic moment ratio v2 changes too.

Lev: From an engineering standpoint, using a machine learning framework like this to find functional forms for those parameters is interesting because it lets you build a bridge between what you can calculate on the lattice and what an effective model predicts Kai Right, so they are essentially trying to teach the machine how the coupling G and that ratio v2 evolve as the magnetic field B increases.

Mira: And it claims that this method successfully identifies continuous parameter functions for G with respect to B and v2 with respect to B, which helps reproduce the inverse magnetic catalysis effect.

Lev: That IMC effect is a tricky one, so showing that a machine learning pipeline can capture it from the lattice data is significant because it means we have a tool to study something hard.

Kai: What matters here is that they embed the Nambu-Jona-Lasinio gap equation directly into their neural network as a physics constrained module.

Mira: That constraint acts like a hard rule, making sure whatever the machine learns still obeys fundamental laws of physics, which is what makes this approach more than just curve fitting data points.

Lev: On the hardware side, if you were trying to run this on real quantum hardware for simulations, you'd need that differentiable module to be very stable because automatic differentiation has its own overhead.

Kai: That's a fair point, Lev; the complexity of that constraint really lands somewhere in terms of computational cost.

Mira: The paper claims they observe that both G and v2 are smoothly suppressed by the external magnetic field across the range from zero to zero point six GeV squared <ref:2603.29269#pg3>.

Kai: Smoothly suppressed, you mean it's not some sudden jump or a sharp cutoff where the physics breaks down?

Paper summary: Lev: It's smooth suppression, which is good for modeling because sudden changes are often more problematic for numerical stability when you're trying to evolve the system over time.

Mira: So what they found in those results is that G decreases from a vacuum value of four point eight zero five times ten to the power minus two GeV minus two to three point five five zero times ten to the power minus two GeV minus two at the highest field strength.

Kai: That's a reduction, which is exactly what we expect when you put a strong magnetic field in because it affects how quarks interact.

Mira: And similarly for the AMM ratio v2, which they found reduced from zero point seven seven times ten to the power minus three GeV minus three to zero point seven three times ten to the power minus three GeV minus three at that same high field strength <ref:2603.29269#pg3>.

Lev: So, if you're trying to run this on a real system, you have these parameters G and v2 that are changing in response to B, which means your simulation setup has to account for that evolution continuously.

Kai: That makes sense; the model isn't just static inputs anymore; it's dynamic based on the field strength.

Mira: The paper also mentions how the chiral condensate shows a complex, temperature dependent dual behavior, capturing that inverse magnetic catalysis effect we've seen in lattice QCD before.

Lev: It’s interesting that the machine learning framework is able to capture this dual behavior without needing all those extra, complex terms explicitly written down in the Lagrangian.

Kai: The authors are essentially saying this differentiable inversion framework can bridge the gap between first principles calculations and these effective models.

Mira: That’s a big statement because it suggests that machine learning isn't just a black box for finding answers; it can actually help us engineer those answers from our existing fundamental data.

Lev: For future work, they plan to extend this pipeline to more complex effective theories, like the Polyakov-loop Nambu-Jona-Lasinio model or even systems with two or three flavors.

Kai: That’s where the real test will be if this framework can handle that added complexity without losing its physical constraints.

Mira: Overall, the implication is that machine learning is a powerful engine for parameter optimization within the context of fundamental physical laws when you have high-dimensional data like lattice QCD results to work with.

Paper summary: Lev: It shows that these complex systems can be tackled by finding optimal functional forms through this inverse engineering approach.

Kai: So, to wrap up on the "Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning" paper, it really demonstrates how we can use machine learning to extract field-dependent parameters from lattice QCD data using a differentiable physics constrained module.

Mira: The authors show that this smoothly suppresses both the coupling constant G and the AMM ratio v2 as you increase the magnetic field B, achieving excellent agreement with lattice data.

Lev: The main limitation they point out is that their reliability might decrease in a regime where the magnetic field strength exceeds two point zero four zero five eight GeV squared because it starts to go beyond the ultraviolet cutoff scale of the model.

Kai: That means if you try to push the magnetic field way past that point, this specific ML pipeline might not be reliable anymore.

Mira: This paper establishes a general paradigm for linking first-principles calculations and phenomenological models by using this differentiable inversion framework.

Kai: It’s about taking complex data, like the quark condensate from lattice QCD, and teaching a neural network how to map that data onto the parameters of an effective model in a physically sound way.

Lev: So, for those of you interested in the actual implementation, this framework shows that you can use gradient descent optimization targeting specific loss functions—like penalizing predictions outside a five percent tolerance zone from the lattice data—to find these optimal functional forms for G and v2.

Kai: That’s a concrete way to see how the learning actually works, not just the abstract idea of "learning."

Mira: And what this means for us is that we have a method to study the response of the QCD vacuum to strong magnetic fields with microscopic insights into how things change.

Lev: It’s a tool for discovery in high-energy physics, leveraging machine learning to handle the complexity where traditional methods might struggle.

Kai: We'll stick with this idea that machine learning is a powerful engine for parameter optimization within the context of fundamental physical laws when you have high-dimensional data like lattice QCD results to work with.

Mira: And we'll keep an eye on how they extend this pipeline to more complex theories in future work.

Conclusion: Kai: So we're wrapping up this look at that paper "Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning".

Mira: Yeah, it’s about using machine learning to figure out how two parameters, the coupling constant G and that anomalous magnetic moment ratio v2, change when you put a strong magnetic field into matter governed by the Nambu-Jona-Lasinio model.

Lev: It boils down to training a neural network to learn those functions based on lattice QCD data as the ground truth.

Kai: The authors basically say they built a way to bridge that gap between those complex first-principles calculations and these simpler effective models.

Mira: They’re using this differentiable framework where the gap equation acts like a hard constraint to make sure whatever the machine learns still follows basic physics laws.

Lev: That constraint is key, but you gotta be careful with how much overhead that adds when you try to run it on real hardware.

Kai: The results show that both G and v2 are smoothly suppressed by the magnetic field as the field strength goes up to zero point six giga-tesla squared.

Mira: They also found that this method successfully captures something called the inverse magnetic catalysis effect which lattice QCD has shown before.

Lev: It's interesting how they managed to encode that complex behavior just by fitting parameters instead of writing out every single term in the Lagrangian manually.

Kai: The main implication is showing that machine learning can be a powerful tool for discovery and parameter optimization when you have high-dimensional data like lattice QCD results to work with.

Mira: It suggests we can use this kind of differentiable inversion framework to study how the QCD vacuum reacts to strong magnetic fields in a much more flexible way.

Lev: We still need to figure out if this specific setup holds up when we start moving into more complicated theories, like those polyakov loop models later on.

Zigeng Ding, Fan Lin, Xinyang Wang

Center for Fundamental Physics, School of Mechanics and Physics, Anhui University of Science and Technology

hep-ph, nucl-th

Submitted: 2026-03-31

Updated: 2026-11-09

Comments: 14pages, 6 figures, The accepted version of EPJC

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning investigates the phase structure of magnetized QCD matter by determining field-dependent parameters of the…

Key concepts

Nambu-Jona-Lasinio (NJL) Model
This is a theoretical model used to describe how quarks interact and form a 'constituent mass' within Quantum Chromodynamics (QCD). It helps physicists understand the properties of strongly interacting matter, especially when external forces like magnetic fields are applied.
Anomalous Magnetic Moment (AMM)
The AMM describes the extra magnetic moment quarks acquire due to quantum effects in a magnetic field. This effect modifies how quarks behave and is crucial for accurately modeling QCD in the presence of strong fields.
Physics-Informed Machine Learning
This method uses machine learning not just to find patterns, but to learn physics directly. The model is constrained by fundamental laws, like the gap equation, ensuring that the learned relationships adhere strictly to known physical principles.

Terminology

Summary

Determining the NJL Coupling and AMM in Magnetized QCD Matter via Machine Learning investigates the phase structure of magnetized QCD matter by determining field-dependent parameters of the Nambu-Jona-Lasinio (NJL) model through a physics-informed machine learning framework The magnetic field smoothly suppresses both G and v2, providing new microscopic insights into the response of the QCD vacuum to strong magnetic fields

Theoretical Framework

The Lagrangian density of the two-flavor Nambu–Jona-Lasinio (NJL) model in the presence of a background magnetic field, including the contribution from the anomalous magnetic moment (AMM) of quarks, is given by L = ψ¯ iγµDµ − mˆ + 1/2 qfκFµνσ µν ψ + G (ψψ¯)2 + (ψiγ ¯ 5 τψ)2 To determine the dynamical quark mass, the mean-field approximation is employed, leading to the constituent quark mass matrix Mˆ = ˆm + σ, where σ = −2G⟨ψψ¯ ⟩ The presence of the AMM modifies the Dirac equation for quarks to iγµDµ − Mˆ + 1/2 κqfσ µνFµν ψ = 0 The thermodynamic potential Ω is obtained following standard procedures, taking the form of a sum over Landau levels

How the training signal is built

The core idea involves employing machine learning as a tool for inverse engineering, using lattice QCD computed magnetic-field-dependent quark condensate data as the “ground truth” and a parameterized (field-dependent) NJL model as the “interpretable function” The neural network is trained to directly “learn” the optimal functional forms for the evolving coupling constant G(eB) and AMM ratio v2(eB) which is defined by κ(eB) = v2(eB)σ2 This involves constructing a differentiable physics-constrained module where the gap equation acts as a “hard constraint” to ensure strict adherence to fundamental physical laws

Where the gains come from

The optimization objective is defined by the total loss function Ltotal(θ) = λbLb + λMLM + λσLσ + λQC (LlowTQC + LhighTQC) This loss function balances data fidelity, empirical constraints, and physical priors The constituent quark mass loss LM penalizes predictions falling outside a ±5% tolerance zone of the lattice data Mlatt The physics-informed regularizer includes losses such as the Chiral Dynamical Quark Mass Monotonicity Loss Lσ, which encodes the physical behavior that σ decreases monotonically with increasing magnetic field at zero temperatures

Numerical Results

Within the investigated range of 0 < eB < 0.6 GeV2, both G(eB) and v2(eB) are observed to be smoothly suppressed by the external magnetic field Specifically, the coupling constant G(eB) decreases from a vacuum value of 4. p. 6

Where the paper says it

6 GeV2 , we observe that the magnetic field smoothly suppresses both parameters. Specifically, the coupling constant G(eB) decreases from a vacuum value of 4.

805 ± 0.012 GeV−2 to 3. p. 6
Where the paper says it

Specifically, the coupling constant G(eB) decreases from a vacuum value of 4. 012 GeV−2 to 3.

550 ± 0.031 GeV−2 at the highest field strength Similarly, the AMM ratio v2 is reduced from 0. p. 6
Where the paper says it

Similarly, the AMM ratio v2 is reduced from 0. 057 GeV−3 to 0.

771 ± 0.057 GeV−3 to 0. p. 6
Where the paper says it

Similarly, the AMM ratio v2 is reduced from 0. 057 GeV−3 to 0.

734 ± 0.059 GeV−3 The chiral condensate displays a complex, temperature-dependent dual behavior, capturing the inverse magnetic catalysis effect observed in lattice QCD

Conclusion and Outlook

The differentiable inversion framework establishes a general paradigm for bridging the gap between first-principles calculations and phenomenological models Future work will focus on extending the pipeline to more sophisticated effective theories, such as the Polyakov-loop NJL (PNJL) model or (2+1)-flavor systems The methodology demonstrates that machine learning is a powerful engine for discovery and parameter optimization within the context of fundamental physical laws

Acknowledgments

The authors thank Zhongmin Niu for insightful discussions and Minghua Wei for early contributions to this work The authors acknowledge support from the National Natural Science Foundation of China (Grant No. 12547138) and the Anhui University of Science and Technology (Grant No. p. 8

Where the paper says it

acknowledges support from the National Natural Science Foundation of China (Grant No. 12547138) and the Anhui University of Science and Technology (Grant No.

2025yjrc0143) X. p. 8
Where the paper says it

12547138) and the Anhui University of Science and Technology (Grant No. 2025yjrc0143).

W. was supported by the Anhui University of Science and Technology under Grant No. p. 8
Where the paper says it

2025yjrc0143). was supported by the Anhui University of Science and Technology under Grant No.

YJ20240001 The paper is available as arXiv:2603. p. 0
Where the paper says it

∗ wangxy@aust. cn, corresponding author arXiv:2603.

29269v1 [hep-ph] 31 Mar 2026 The corresponding author is wangxy@aust. p. 0
Where the paper says it

cn, corresponding author arXiv:2603. 29269v1 [hep-ph] 31 Mar 2026

edu.cn

References

[1] S. P. Klevansky and Richard H. p. 8

Where the paper says it

YJ20240001. Klevansky and Richard H.

Lemmer. Chiral symmetry restoration in the Nambu-Jona-Lasinio model with a constant electromagnetic field. p. 8
Where the paper says it

Klevansky and Richard H. Chiral symmetry restoration in the Nambu-Jona-Lasinio model with a constant electromagnetic field.

Phys. Rev. D, 39:3478–3489, 1989

[2] K. G. Klimenko. Three-dimensional Gross-Neveu model in an external magnetic field. p. 8

Where the paper says it

D, 39:3478–3489, 1989. Three-dimensional Gross-Neveu model in an external magnetic field.

Teor. Mat. Fiz., 89:211–221, 1991

[3] V. P. Gusynin, V. p. 8

Where the paper says it

, 89:211–221, 1991. Gusynin, V.

A. Miransky, and I. p. 8
Where the paper says it

Gusynin, V. Miransky, and I.

A. Shovkovy. Dimensional reduction and catalysis of dynamical symmetry breaking by a magnetic field. p. 8
Where the paper says it

Miransky, and I. Dimensional reduction and catalysis of dynamical symmetry breaking by a magnetic field.

Nucl. Phys. B, 462:249–290, 1996

[4] Igor A. Shovkovy. Magnetic Catalysis: A Review. p. 8

Where the paper says it

[4] Igor A. Magnetic Catalysis: A Review.

Lect. Notes Phys. p. 8
Where the paper says it

Magnetic Catalysis: A Review. Notes Phys.

, 871:13–49, 2013

[5] G. S. Bali, F. Bruckmann, G. p. 8

Where the paper says it

, 871:13–49, 2013. Bruckmann, G.

p. 8
Where the paper says it

, 871:13–49, 2013. Bruckmann, G.

Endrodi, Z. p. 8
Where the paper says it

Bruckmann, G. Endrodi, Z.

p. 8
Where the paper says it

Bruckmann, G. Endrodi, Z.

Fodor, S. D. Katz, S. Krieg, A. Schafer, and K. p. 8
Where the paper says it

Endrodi, Z. Schafer, and K.

K Szabo The QCD phase diagram for external magnetic fields JHEP 02:044 2012

[6] G. S. Bali, F. Bruckmann, G. Endrodi, Z. Fodor, S. D Katz, and A Schafer QCD quark condensate in external magnetic fields Phys. p. 8

Where the paper says it

Katz, and A. QCD quark condensate in external magnetic fields.

Rev. D 86:071502 2012

[7] H. T Ding, S. T Li, J H Liu, and X D Wang Chiral condensates and screening masses of neutral pseudoscalar mesons in thermomagnetic QCD medium Phys. p. 8

Where the paper says it

Liu, and X. Chiral condensates and screening masses of neutral pseudoscalar mesons in thermomagnetic QCD medium.

Rev. D 105(3):034514 2022

[8] Yoichiro Nambu and G Jona-Lasinio Dynamical Model of Elementary Particles Based on an Analogy with Superconductivity. Phys. Rev., 122:345–358, 1961

[9] Yoichiro Nambu and G Jona-Lasinio Dynamical model of elementary particles based on an analogy with superconductivity Phys. Rev., 124:246–254, 1961

[10] R L S Farias, K P Gomes, G I Krein, and M B Pinto Importance of asymptotic freedom for the pseudocritical temperature in magnetized quark matter Phys. Rev. C 90(2):025203 2014

[11] M Ferreira, P Costa, O Lourenço, T Frederico, and C Providência Inverse magnetic catalysis in the (2+1)-flavor Nambu-Jona-Lasinio and Polyakov-Nambu-Jona-Lasinio models Phys. Rev. D 89(11):116011 2014

[12] Hao Liu, Lang Yu, Maxim Chernodub, and Mei Huang Possible formation of high temperature superconductor at an early stage of heavy-ion collisions Phys. Rev. D 94(11):113006 2016

[13] Bing-kai Sheng, Xinyang Wang, and Lang Yu Impacts of inverse magnetic catalysis on screening masses of neutral pions and sigma mesons in hot and magnetized quark matter Phys. Rev. D 105(3):034003 2022

[14] Sidney S Avancini, Maximo Coppola, Norberto N Scoccola, and Joana C Sodré Light pseudoscalar meson masses under strong magnetic fields within the SU(3) Nambu–Jona-Lasinio model Phys. Rev. D 104(9):094040 2021

[15] Shijun Mao and Yvming Tian Pion superfluid phase transition under an external magnetic field including the inverse magnetic catalysis effect Phys. Rev. D 106(9):094017 2022

[16] Shijun Mao Correlations and fluctuations in a magnetized PNJL model with and without the inverse magnetic catalysis effect Chin. Phys. C, 49(6):063106 2025

[17] Shijun Mao and Shuai Yang Correlations and fluctuations in a magnetized three-flavor PNJL model with and without inverse magnetic catalysis effect Phys. Rev. D 112(1):014026 2025

[18] E J Ferrer, V de la Incera, and X J Wen Quark Antiscreening at Strong Magnetic Field and Inverse Magnetic Catalysis Phys. Rev. D 91(5):054006 2015

[19] Efrain J Ferrer and Vivian de la Incera Dynamically Generated Anomalous Magnetic Moment in Massless QED Nucl. Phys. B 824:217–238, 2010

[20] Efrain J Ferrer, Vivian de la Incera, Israel Portillo, and Matthew Quiroz New look at the QCD ground state in a magnetic field Phys. Rev. D 89(8):085034 2014

[21] Pedro J A Bicudo, J Emilio F T Ribeiro, and Rui Fernandes The Anomalous magnetic moment of quarks Phys. Rev. C 59:1107–1112, 1999

[22] Lei Chang, Yu-Xin Liu, and Craig D Roberts Dressed-quark anomalous magnetic moments Phys. Rev. Lett., 106:072001 2011

[34] Fan Lin and Mei Huang Magnetic correction to the anomalous magnetic moment of electrons Commun. Theor. Phys., 74(5):055202 2022

[35] Fan Lin, Kun Xu, and Mei Huang Magnetism of QCD matter and the pion mass from tensor-type spin polarization and the anomalous magnetic moment of quarks Phys. Rev. D 106(1):016005 2022

[36] Mamiya Kawaguchi and Mei Huang Restriction on the form of the quark anomalous magnetic moment from lattice QCD results Chin. Phys. C, 47(6):064103 2023

[37] De-Xian Wei and Li-Juan Zhou Quark condensate and magnetic moment in a strong magnetic field Int. J. Mod. Phys. E, 32(05):2350024 2023

[38] Mamiya Kawaguchi, Irfan Siddique, and Mei Huang Effect of quark anomalous magnetic moment on neutral dense quark matter under magnetic field Eur. Phys. J. C, 85(3):246 2025

[39] Jürgen Schmidhuber Deep learning in neural networks: An overview Neural Networks, 61:85–117, 2015

[40] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton Deep learning. Nature, 521:436–444, 2015

[41] Amber Boehnlein et al Colloquium Machine learning in nuclear physics Rev. Mod. Phys., 94(3):031003 2022

[42] Andrew J Larkoski, Ian Moult, and Benjamin Nachman Jet substructure at the Large Hadron Collider A review of recent advances in theory and machine learning Phys. Rept., 841:1–63 2020

[43] Pierre Baldi, Kyle Cranmer, Taylor Faucett, Peter Sadowski, and Daniel Whiteson Parameterized neural networks for high-energy physics Eur. Phys. J. C, 76(5):235 2016

[44] Yi-Lun Du, Daniel Pablos, and Konrad Tywoniuk Deep learning jet modifications in heavy-ion collisions JHEP 21:206 2020

[45] Long-Gang Pang, Kai Zhou, Nan Su, Hannah Petersen, Horst Stöcker, and Xin-Nian Wang An equation-of-state-meter of quantum chromodynamics transition from deep learning Nature Commun., 9(1):210 2018

[46] Long-Gang Pang, Kai Zhou, and Xin-Nian Wang Interpretable deep learning for nuclear deformation in heavy ion collisions arXiv: 1906.06429 [nucl-th] 6 2019

[47] Yi-Lun Du, Kai Zhou, Jan Steinheimer, Long-Gang Pang, Anton Motornenko, Hong-Shi Zong, Xin-Nian Wang and Horst Stöcker Identifying the nature of the QCD transition in relativistic collision of heavy nuclei with deep learning Eur. Phys. J. C 80(6):516 2020

[48] Wan-Bing He, Yu-Gang Ma, Long-Gang Pang, Hui-Chao Song, and Kai Zhou High-energy nuclear physics meets machine learning Nucl. Sci. Tech., 34(6):88 2023

[49] Kai Zhou, Lingxiao Wang, Long-Gang Pang, and Shuzhe Shi Exploring QCD matter in extreme conditions with Machine Learning Prog. Part. Nucl. Phys., 135:104084 2024

[50] Yu-Gang Ma, Long-Gang Pang, Rui Wang, and Kai Zhou Phase Transition Study Meets Machine Learning Chin. Phys. Lett., 40(12):122101 2023

[51] Yongjia Wang and Qingfeng Li Machine learning transforms the inference of the nuclear equation of state Front. Phys (Beijing), 18(6):64402 2023

[52] Wanbing He, Qingfeng Li, Yugang Ma, Zhongming Niu, Junchen Pei, and Yingxun Zhang Machine learning in nuclear physics at low and intermediate energies Sci. China Phys. p. 9

Where the paper says it

Machine learning in nuclear physics at low and intermediate energies. China Phys.

Mech. Astron., 66(8):282001 2023

[53] Rui Wang, Yu-Gang Ma, R Wada, Lie-Wen Chen, Wan-Bing He, Huan-Ling Liu, and Kai-Jia Sun Nuclear liquid-gas phase transition with machine learning Phys. Rev. Res., 2(4):043202 2020

[54] Yu-Jing Huang, Zhu Meng, Long-Gang Pang, and Xin-Nian Wang A Novel Deep Learning Method for Detecting NucleonNucleon Correlations arXiv: 2504.00790 [nucl-th] 4 2025

[55] Michael Buballa NJL model analysis of quark matter at large density Phys. Rept., 407:205–376, 2005

[56] Kai Xue, Xiaozhu Yu, and Xinyang Wang Do we need to use regularization for the thermal part in the NJL model? Chin. Phys. C, 46(5):013103 2022

Key Hyperparameters

The configuration of hyperparameters and weighting coefficients for the multi-objective loss function is detailed in Table II The weighting coefficients λi are empirically determined to implement a hierarchical optimization priority, assigning significantly higher magnitudes to the physics-informed weights (λσ, λQ) and the constituent quark mass constraint (λM) relative to the temperature weight λT The model is trained using a two-stage optimization strategy, starting with pre-training solely on data-driven terms before activating the full loss function for joint optimization This ensures that the interplay between high-weight physical constraints and the data-fidelity term reaches a global minimum

Final Findings

The learned field dependence allows the NJL model to naturally resolve the long-standing discrepancy regarding the inverse magnetic catalysis effect at high temperatures The framework achieves excellent agreement with lattice data, with a maximum discrepancy of less than 1% in Fig. 5 The reliability of the model may decrease in the regime where eB > Λ2 ≈ 0. p. 7

Where the paper says it

75 GeV−3 × σ 2 . In the end, it is important to note that the reliability of the model may decrease in the regime where |eB| > Λ 2 ≈ 0.

4058 GeV2, as the field strength begins to exceed the ultraviolet cutoff scale The authors estimate that the quark AMM term is in the order of magnitude κ ≈ 0. p. 7
Where the paper says it

In the end, it is important to note that the reliability of the model may decrease in the regime where |eB| > Λ 2 ≈ 0. 4058 GeV2 , as the field strength begins to exceed the ultraviolet cutoff scale.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved parameter extraction for effective models by using a physics-informed machine learning framework that embeds the NJL gap equation as a differentiable physics-constrained module. This allows neural networks to identify continuous parameter functions that accurately reproduce the inverse magnetic catalysis (IMC) effect by inverting for G(eB) and v2(eB).

  2. Enhanced capability of deep learning models to discover physical laws from high-dimensional data by employing an inversion framework where "lattice QCD computed quark condensate data as the “ground truth” and a parameterized (field-dependent) NJL model as the “interpretable function.” This bridges the gap between effective models and lattice data.

  3. Ability to perform parameter-inversion problems for unknown functions like G(eB) and v2(eB) by utilizing a two fully connected neural networks (FCNNs) with identical architectures, each dedicated to parameterizing one function. This allows the system to learn only the deviations from fixed baselines relative to reference values.

  4. Robust optimization of physical parameters using a multi-objective loss function that reconciles model predictions with data and theory by minimizing Ltotal(theta). This loss includes terms like Normalized Critical Temperature Boundary Loss and Constituent Quark Mass Loss, ensuring both empirical fidelity and physical consistency.

  5. Implementation of a two-stage optimization strategy, beginning with pre-training solely on data-driven terms and moving to joint optimization with full loss function activation. This ensures a robust convergence that honors both empirical observations and theoretical requirements by allowing the network to capture foundational dependencies before applying complex physical constraints.

  6. Direct prediction of key QCD observables like the constituent quark mass Mpred(eBi) at zero temperature and the normalized critical temperature ˆy(eBi) = Tc(eBi)/Tc(eB = 0) for various magnetic field strengths. This enables rapid evaluation of how the model parameters affect physical properties across a range of magnetic field strengths.

  7. Identification of the physical mechanisms driving phase transitions by analyzing the behavior of observables, such as "The chiral condensate displays a complex, temperature-dependent dual behavior: at low temperatures, the magnetic catalysis effect dominates as the field enhances the condensate; however, near the transition region, the field-dependent parameters trigger a suppression of the condensate."

Abstract

In this study, we investigate the phase structure of magnetized QCD matter by determining the field-dependent parameters of the Nambu-Jona-Lasinio (NJL) model through a physics-informed machine learning framework. Specifically, we focus on extracting the optimal functional forms for the running coupling constant G(eB) and the quark anomalous magnetic moment (AMM) ratio v 2(eB), using lattice QCD computation on the chiral transition and condensate behavior. By embedding the NJL gap equation as a differentiable physics-constrained module, our neural network pipeline identifies continuous parameter functions that accurately reproduce the inverse magnetic catalysis (IMC) effect. Our results demonstrate that the magnetic field smoothly suppresses both G and v 2. This approach not only bridges the gap between effective models and lattice data but also provides new microscopic insights into the response of the QCD vacuum to strong magnetic fields. As an independent test, the neutral-pion mass is calculated with the mean learned parameters and reproduces the decreasing trend of recent lattice results without a pion-data refit.

Sources

Related papers