Machine Learning Etudes in Conformal Field Theories

Heng-Yu Chen, Shailesh Lal, Yang‐Hui He, M. Zaid Zaz

TL;DR

This paper explores the application of machine learning, specifically neural networks, to various problems in Conformal Field Theory (CFT). The authors demonstrate that neural networks can effectively learn and identify fundamental symmetries and physical data within CFTs, offering a powerful new tool for theoretical physics research.

What Problem Does This Paper Address?

Conformal Field Theories (CFTs) are crucial for understanding critical phenomena in physics, particularly phase transitions. While significant progress has been made in analyzing CFTs, especially in two dimensions, extracting precise physical data (like spectrum and operator product expansion coefficients) in higher dimensions remains computationally challenging. This paper investigates whether machine learning techniques can automate and simplify the analysis of CFT data, potentially uncovering hidden structures and accelerating research in this complex field. The authors aim to build a toolbox of machine learning methods applicable to CFT problems, demonstrating their learnability and utility.

Methods

The core methodology employed is machine learning, primarily using feed-forward neural networks. The authors apply these networks to several tasks:

  • Classification: Distinguishing between functions with different symmetries (even/odd, scale vs. conformal invariance, crossing symmetry), and classifying operators based on their discrete symmetries (Z2, Z3).
  • Regression: Predicting specific coefficients (Fourier coefficients, OPE coefficients) and learning the spectrum of operators (predicting operator twists).

The paper also utilizes Principal Component Analysis (PCA) to visualize high-dimensional data and gain insights into the underlying structures that neural networks learn.

Datasets

The paper uses several datasets, often generated synthetically for specific classification and regression tasks:

  • Even/Odd Functions: Datasets generated from functions with defined even/odd properties.
  • Symmetry Classification: Point clouds representing functions with plane reflection, antipodal reflection, or null symmetries.
  • Scale vs. Conformal Invariance: Datasets based on three-point and two-point correlation functions, along with conformal dimensions.
  • Crossing Symmetric Four-Point Functions: Datasets generated to distinguish crossing symmetric functions from non-symmetric ones.
  • Conformal Block Expansion: Datasets derived from truncated conformal block expansions, focusing on the presence or absence of specific terms.
  • 3D Ising Model Data:
    • Z2 Symmetry Classification: OPE coefficients from the 3D Ising model, transformed to be in the O(1) range.
  • Z3 Symmetry Classification: Synthetic data mimicking OPE coefficients for a Z3 symmetric CFT.
  • Spectrum Learning: Numerical data for operator twists and spins from the 3D Ising model.

Key Insights and Ideas

  • Learnability of CFT Symmetries: Neural networks can effectively distinguish between scale and conformal invariance, and recognize crossing symmetry in four-point functions with high accuracy, even without explicit knowledge of the underlying physics.
  • Decomposition of Complex Functions: Neural networks can identify the presence or absence of specific terms (like conformal blocks or Fourier modes) in complex expansions, suggesting a deep understanding of the function's structure.
  • OPE Coefficient Prediction: Neural networks can learn to predict OPE coefficients, a crucial piece of physical data in CFT, with remarkable accuracy.
  • Operator Classification by Symmetry: Machine learning can classify operators based on their behavior under discrete symmetries (like Z2 and Z3), mirroring fundamental properties of the theory.
  • Spectrum Learning from Data: Neural networks can learn the relationship between operator spin and twist from numerical data, offering a powerful regression tool for extracting spectral information.
  • PCA as a Complementary Tool: PCA helps visualize the high-dimensional data that neural networks process, revealing that even when functions appear similar, their underlying structures are distinct and learnable by neural networks.

Key Results

  • Classification Accuracy: Achieved accuracies of nearly 100% for distinguishing scale vs. conformal invariance, recognizing crossing symmetry, and classifying Z2 symmetry in the 3D Ising model.
  • Regression Accuracy: Obtained R-squared values close to 1 (e.g., 0.9999 for Fourier coefficients, 0.9988 for OPE coefficients, and 0.9966 for 3D Ising model operator twists).
  • Conformal Block Learning: Neural networks could distinguish functions with and without specific conformal blocks with 96-97% accuracy.
  • 3D Ising Model Spectrum: Neural networks successfully predicted operator twists with high R-squared values (0.9804 for [σσ]o and 0.9966/0.9926 for [σε]o).

Limitations and Future Work

  • Data Transformation: For the 3D Ising model OPE data, a transformation (logarithm) was necessary to bring the numerical values into a range suitable for neural network training. This suggests that the range and distribution of data can significantly impact performance.
  • Network Architecture Sensitivity: While the paper uses feed-forward networks, the choice of activation function (tanh) seemed particularly effective for certain regression tasks, hinting at an interplay between function properties and network architecture.
  • Future Work: The authors suggest extending these methods to higher-rank symmetry groups (Z4, Z2 × Z2, SN × Z2) and exploring their application to CFTs studied with conformal bootstrap in higher dimensions. They also express interest in applying these techniques to larger datasets and more complex conformal bootstrap problems.

Summary updated

Abstract

We demonstrate that various aspects of Conformal Field Theory are amenable to machine learning. Relatively modest feed-forward neural networks are able to distinguish between scale and conformal invariance of a three-point function and identify a crossing-symmetric four-point function to nearly a hundred percent accuracy. Furthermore, neural networks are also able to identify conformal blocks appearing in a putative CFT four-point function and predict the values of the corresponding OPE coefficients. Neural networks also successfully classify primary operators by their quantum numbers under discrete symmetries in the CFT from examining OPE data. We also demonstrate that neural networks are able to learn the available OPE data for scalar correlation function in the 3d Ising model and predict the twists of higher-spin operators that appear in scalar OPE channels by regression.
BibTeX
@article{chen2020machine,
  author = {Heng-Yu Chen and Shailesh Lal and Yang‐Hui He and M. Zaid Zaz},
  title = {{Machine Learning Etudes in Conformal Field Theories}},
  journal = {Int. J. Data Sci. Math. sci},
  year = {2020},
}
Figure