Learning geometries beyond asymptotic AdS

Ran Cheng, Shao-Feng Wu, Zhuo-Yu XianarXiv 2025 · DOI · arXiv

This paper, recently published by researchers from Shanghai University, Yangzhou University, and Freie Universität Berlin, explores a frontier in theoretical physics: using machine learning to "see" the hidden dimensions of space-time predicted by the holographic principle.

TL;DR

The authors present a data-driven method using Neural ODEs to reconstruct higher-dimensional "bulk" geometries from boundary quantum field theory data. Unlike previous methods, this approach works even when the system isn't perfectly conformal (non-AdS), successfully identifying a bulk dual for the famous Sachdev-Ye-Kitaev (SYK) model.

What Problem Does This Paper Address?

The AdS/CFT correspondence (or holography) is a powerful tool suggesting that a gravity theory in a higher-dimensional "bulk" space is equivalent to a quantum field theory on its "boundary." Traditionally, this requires the boundary to be a Conformal Field Theory (CFT), which corresponds to a bulk that looks like Anti-de Sitter (AdS) space at infinity.

However, real-world systems—like high-temperature superconductors or the quark-gluon plasma—are rarely perfectly conformal. Furthermore, experimental data is usually limited to a finite frequency window, making it impossible to reconstruct the "asymptotic" (infinite distance) structure of the bulk. This paper tackles the challenge of bulk reconstruction when the "standard" AdS rules don't strictly apply.

Methods

The researchers bridge the gap between physics and AI using two main pillars:

  1. Holographic Wilsonian Renormalization Group (HWRG): In physics, the "radial" direction of the bulk is often interpreted as the energy scale of the boundary theory. The authors use HWRG to derive a radial flow equation—a first-order differential equation that describes how the system's response (Green functions) changes as you move into the bulk.
  2. Neural ODEs: They transform this flow equation into a Neural Ordinary Differential Equation. Instead of a standard deep neural network with discrete layers, a Neural ODE models continuous dynamics. Here, the unknown metric of the space-time (the "shape" of the bulk) is parameterized by a neural network and learned by minimizing the difference between the model's output and boundary data.
  3. Loss Function: They use a loss function based on the Kullback–Leibler (KL) divergence, which measures the information loss between the predicted and actual probability distributions of the boundary sources.

Datasets

The authors validated their model using synthetic data generated from known theoretical frameworks:

  • CFT₁: A 1D conformal field theory (the baseline).
  • Lifshitz Spacetime: A geometry with scaling symmetry that differs from AdS.
  • Hyperscaling-Violating Geometries: Complex metrics used to model certain condensed matter phases.
  • The SYK Model: A landmark quantum mechanical model of Majorana fermions. They used the analytical large-$q$ expression for the SYK two-point function as their primary "test subject" for non-conformal reconstruction.

Key Insights and Ideas

  • The Radial Direction as a Flow: The most elegant insight is that the radial direction in holography is mathematically equivalent to the "time" or "depth" in a Neural ODE. This makes ODE-based machine learning a natural fit for holographic reconstruction.
  • The "Critical Curve" in SYK: When applying the model to the SYK model (which is nearly but not perfectly conformal), the authors didn't find just one bulk. Instead, they discovered a critical curve in the parameter space of temperature ($T$) and radial cutoff ($\epsilon$).
  • Holographic Dual of Coupling: Along this critical curve, the learned metric looks exactly like an AdS₂ black hole with a finite cutoff. This allowed the authors to establish a linear relationship between the SYK coupling constant ($v$) and a bulk parameter ($\tilde{v}$), effectively "mapping" the strength of quantum interactions to a geometric property.

Key Results

  • High Precision: The Neural ODE successfully reconstructed AdS, Lifshitz, and hyperscaling-violating metrics with extremely low error (Loss $\approx 10^{-10}$).
  • SYK Mapping: For the SYK model, they derived an approximate analytical expression for the critical curve: $\epsilon = 1 / (1 + 2\pi \bar{T} / \tilde{v})$.
  • Robustness: Along the critical curve, the learned metric was found to be insensitive to the initial random weights of the neural network, suggesting that the model is capturing a fundamental physical truth rather than a numerical artifact.

Limitations and Future Work

While a significant step forward, the authors acknowledge several areas for expansion:

  • Spatial Dependence: The current study focuses on Green functions at zero wavevector ($k=0$). Future versions will need to incorporate spatial dependence to constrain more complex geometries.
  • Fermionic Data: The paper uses scalar operators; incorporating fermionic data (like ARPES spectrums) would bring the model closer to real-world condensed matter experiments.
  • High-Frequency Modes: The study focused on low-frequency data. To capture high-frequency physics, the model would need to account for "double-trace" couplings ($h$) in the Wilsonian action, which were simplified in this work.
  • Real-World Data: The ultimate goal is to apply this "AdS/DL correspondence" to experimental data from QCD or "strange metals" to uncover the geometries of the real world.

Summary updated

Abstract

We present a data-driven method for holographic bulk reconstruction that works even when the spacetime is not asymptotically AdS. Given the data of boundary Green functions within a finite frequency window, we iteratively adjust a bulk metric with a finite radial cutoff until its holographic Green functions reproduce the boundary data. Based on the holographic Wilsonian renormalization group for the Klein-Gordon equation in an undetermined curve space, we construct a radial flow equation and transform it into a Neural ODE, which is an infinite-depth neural network for modeling continuous dynamics. Assuming the double-trace coupling $h$ in the Wilsonian action is real, we demonstrate that the Neural ODE can effectively learn the metrics with AdS, Lifshitz, and hyperscaling violated asymptotics. In particular, we apply the algorithm to the Sachdev-Ye-Kitaev (SYK) model which slightly deviates from the conformal limit. In the hyperparameter space spanned by the rescaled temperature $\bar{T}$ and the radial cutoff $ε$, we identify a critical curve along which the learned metric is close to AdS$_2$ black hole with finite cutoff. We derive an approximate analytical expression for this curve, from which an effective bulk dual of the SYK coupling $v$ is established. Our work provides a promising way for using machine learning to depict the novel bulk geometry dual to the non-conformal boundary systems in the real world.
BibTeX
@misc{cheng2025learning,
  author = {Ran Cheng and Shao-Feng Wu and Zhuo-Yu Xian},
  title = {{Learning geometries beyond asymptotic AdS}},
  year = {2025},
  eprint = {2508.05808},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.05808},
  howpublished = {arXiv},
}

Loading related content...

Figure