HoloNet: Toward a Unified Einstein-Maxwell-Dilaton Framework of QCD

We propose HoloNet, a neural-network framework that unifies lattice QCD(LQCD) thermodynamics and holographic Einstein-Maxwell-Dilaton (EMD) theory within a data-to-holography pipeline. Instead of assuming specific functional forms, HoloNet learns the metric profile $A(z)$ and the gauge-dilaton coupling $f(z)$ directly from 2+1-flavor LQCD data at $μ=0$. These learned functions are embedded into the EMD equations, enabling the model to reproduce the lattice equation of state and baryon number fluctuations with high fidelity. Once trained, HoloNet provides a fully data-driven holographic description of QCD that extends naturally to finite density, allowing us to map the phase diagram and estimate the location of the critical end point (CEP). The reconstructed potential $V(ϕ)$ and coupling $f(ϕ)$ agree quantitatively with those obtained from holographic renormalization, demonstrating that HoloNet can consistently bridge different holographic models.

💡 Research Summary

This paper introduces “HoloNet,” a novel neural network-based framework designed to construct a holographic model of Quantum Chromodynamics (QCD) directly from first-principles lattice QCD (LQCD) data. The core innovation lies in bypassing the traditional, often arbitrary, assumptions about the functional forms of key functions in the bottom-up holographic Einstein-Maxwell-Dilaton (EMD) model. Instead of pre-defining the metric warp factor A(z) or the gauge-dilaton coupling f(φ), HoloNet employs deep neural networks to learn these functions directly from 2+1 flavor LQCD data at zero baryon chemical potential (μ=0).

The theoretical foundation is the 5D EMD action, whose equations of motion are cast into an integral form. This formulation allows the thermodynamic observables—such as temperature, entropy density, and baryon number susceptibility (χ_B^2)—to be expressed as functionals of A(z) and f(z). HoloNet ingeniously embeds these integral equations as a fixed computational graph within a neural network architecture. Two trainable sub-networks output the values of A(z) and f(z) at any given holographic coordinate z. These outputs are then fed into the fixed graph to compute all physical quantities. Key physics constraints, like the asymptotic AdS boundary condition, the Stefan-Boltzmann limit, and the positivity of f(z), are hard-coded into the network design via specific activation functions (e.g., softplus) and included in the loss function as regularization terms.

The training process is conducted in two stages for efficiency and clarity. First, the sub-network for A(z) is optimized using LQCD data for the equation of state (EoS), which depends solely on the metric. Subsequently, with A(z) fixed, the sub-network for f(z) is trained using data for the baryon number susceptibility χ_B^2, which encodes information about the system’s response to a chemical potential. Once trained on μ=0 data, the model possesses a fully determined f(z), enabling self-consistent predictions at finite μ without any further assumptions. This allows HoloNet to extrapolate into the finite-density regime, map the QCD phase diagram, and estimate the location of the critical endpoint (CEP).

The authors demonstrate that HoloNet successfully reproduces the LQCD EoS and χ_B^2 with high fidelity. Furthermore, they show that the dilaton potential V(φ) and coupling f(φ) reconstructed from the learned A(z) and f(z) agree quantitatively with those obtained from independent holographic renormalization methods. This agreement validates the consistency of the data-driven approach. By significantly reducing the human-introduced priors inherent in traditional holographic model building, HoloNet presents a unified, less biased framework that can bridge different holographic constructions and provide a robust, data-informed description of strongly coupled QCD thermodynamics.