Abstract
The neural ordinary differential equation (Neural ODE) is a novel machine learning architecture whose weights are smooth functions of the continuous depth. We apply the Neural ODE to holographic QCD by regarding the weight functions as a bulk metric, and train the machine with lattice QCD data of chiral condensate at finite temperature. The machine finds consistent bulk geometry at various values of temperature and discovers the emergent black hole horizon in the holographic bulk automatically. The holographic Wilson loops calculated with the emergent machine-learned bulk spacetime have consistent temperature dependence of confinement and Debye-screening behavior. In machine learning models with physically interpretable weights, the Neural ODE frees us from discretization artifact leading to difficult ingenuity of hyperparameters, and improves numerical accuracy to make the model more trustworthy.
Introduction
In theoretical physics, the AdS/CFT correspondence (or holography) suggests that a complex quantum theory on a boundary, such as Quantum Chromodynamics (QCD), can be described by a simpler gravitational theory in a higher-dimensional "bulk" spacetime. This duality provides a powerful tool for studying the strong nuclear force, which governs the interactions of quarks and gluons.
Previous attempts to use machine learning to "learn" this bulk geometry from boundary data relied on standard deep neural networks. In these models, each layer represented a discrete slice of space. However, physical spacetime is inherently continuous and smooth. Discrete networks often suffer from "discretization artifacts," requiring researchers to implement complex mathematical penalties to maintain physical consistency. Moreover, simple integration methods used in discrete models, like the Euler method, lack the accuracy needed to handle the extreme conditions near a black hole horizon, where the geometry becomes singular.
Methodology
To overcome these limitations, researchers have successfully applied the Neural Ordinary Differential Equation (Neural ODE) framework to holographic QCD. In this setup, the network's depth is treated as the continuous radial coordinate ($\eta$) of the holographic bulk. The trainable weights of the network are interpreted as the metric function $h(\eta)$, which defines the curvature of the learned spacetime.
The model is trained using Lattice QCD data from the RBC-Bielefeld collaboration, specifically focusing on the chiral condensate as a function of quark mass at various temperatures (0.188 to 0.208 GeV). The equations of motion for a scalar field in curved spacetime are mapped directly onto the ODE that the neural network solves. By using an advanced adaptive solver ("dopri5"), the model maintains high numerical accuracy even as the geometry approaches a singularity. The training process acts as a binary classifier, distinguishing physical observations from random noise to find the spacetime geometry that best explains the data.
Key Insights and Results
The most striking finding is the automatic discovery of the black hole horizon. Even when initialized with a simple Taylor series that does not assume a horizon, the model naturally evolves toward a divergent metric function at the location where a black hole should physically exist. This emergent spacetime is highly interpretable; unlike "black box" AI models, the weights here have direct physical meaning as the metric of the universe.
Furthermore, the model suggests that the bulk metric $h(\eta)$ remains largely independent of temperature, while the temperature dependence of the boundary physics is primarily captured by the scalar field's coupling constant. This separation simplifies the physical description and highlights the robustness of the learned geometry.
Physical Validation
To validate the learned geometry, the authors calculated Wilson loops, which describe the potential $V(d)$ between a quark and an antiquark. The results correctly reproduced three distinct physical phases:
- Coulombic: Short-distance attraction.
- Confinement: A linear potential at medium distances.
- Debye Screening: A flattening of the potential at high temperatures, consistent with the behavior of a Quark-Gluon Plasma.
The adaptive ODE solver successfully handled the coordinate singularity at the black hole horizon, a task that typically causes simpler discrete models to fail. This demonstrates the superior numerical stability and physical fidelity of the Neural ODE approach.
Conclusion and Future Work
By treating neural network depth as a continuous spatial dimension, the Neural ODE framework provides a more natural and accurate way to model holographic spacetime. It eliminates the need for hand-tuned smoothness regularizations and offers a more trustworthy model for theoretical physics.
Future research aims to scale this model to more complex Einstein-dilaton systems, which would provide a more complete gravitational description of QCD. Additionally, the authors plan to train the model to find all metric components independently, further reducing the reliance on assumed functional forms and moving toward a fully emergent description of the holographic bulk.
Summary updated