Source-linked AI summary
Machine learning phases of matter
Juan Carrasquilla, Roger G. Melko
TL;DR
The paper asks whether supervised neural networks can identify diverse phases and phase transitions directly from raw configurations without being given the Hamiltonian or interaction locality. It trains feed-forward and convolutional networks on sampled states and finds that they classify conventional, Coulomb, and topological phases, including perfect classification of T = 0 and T = ∞ states in the reported test.
Problem
The paper investigates whether neural networks can identify phases, including phases without conventional order parameters, from raw state configurations without Hamiltonian or interaction-locality knowledge.
Method
The paper uses supervised neural networks trained on raw configurations, including a convolutional architecture for topological phases, to classify sampled condensed-matter states.
Results
100% accuracy was obtained when the CCN classified T = 0 and T = ∞ states.
Takeaways & Limitations
The results show that the network’s classification relies on satisfied local constraints rather than topological order for the examined Ising lattice gauge states.
Abstract
from arXiv · showhide
Neural networks can be used to identify phases and phase transitions in condensed matter systems via supervised machine learning. Readily programmable through modern software libraries, we show that a standard feed-forward neural network can be trained to detect multiple types of order parameter directly from raw state configurations sampled with Monte Carlo. In addition, they can detect highly non-trivial states such as Coulomb phases, and if modified to a convolutional neural network, topological phases with no conventional order parameter. We show that this classification occurs within the neural network without knowledge of the Hamiltonian or even the general locality of interactions. These results demonstrate the power of machine learning as a basic research tool in the field of condensed matter and statistical physics.
Appendix A: Details of the toy model
The toy model encodes Ising low- and high-temperature phases through three hidden perceptrons whose activations depend on magnetization thresholds. After training, the network’s hidden-layer arguments become approximately linear in magnetization, supporting magnetization encoding in the hidden layer.
- Toy-model architecture: The toy model represents Ising phases with three hidden perceptrons and recombines their activations in the output layer for classification.The hidden units distinguish polarized and unpolarized configurations, while the output units encode low- and high-temperature states.
- Toy-model architecture: ϵ is the magnetization threshold determining whether a state is treated as polarized.The model assumes 0 < ϵ < 1, with polarization identified when magnetization exceeds positive or negative threshold values.
- Toy-model architecture: The first two hidden neurons activate for predominantly polarized states, while the third activates when m(x) > −ϵ.In the limit 0 < ϵ ≪ 1, the third neuron also activates for nearly all states except sufficiently down-polarized configurations.
- Output classification: The network’s output assigns OLow-T = 1 when spins are mostly polarized and OHigh-T = 1 for the corresponding high-temperature activation pattern.The output layer recombines hidden-neuron outcomes to produce the desired phase classification.
- Training behavior: After training, the components of W^t x + b^t become approximately linear functions of magnetization m(x).The trained behavior agrees with the analytical toy-model assumptions and supports encoding magnetization in the hidden layer.
Appendix B: Visualizing the action of a neural network on the Ising ferromagnet
The paper visualizes how the neural network separates Ising configurations using low-dimensional embeddings and hidden or output-layer responses. In the t-SNE representation, the two ordered low-temperature states form separate regions from the high-temperature paramagnetic region.
- t-SNE visualization: t-SNE embeds high-dimensional Ising configurations into two or three dimensions while preserving proximity between nearby points.The visualization is used to inspect the structure of the training data in a low-dimensional space.
- t-SNE visualization: The t-SNE map contains two low-temperature blue regions corresponding to spins polarized up or down.These regions represent the two ordered Ising states.
- t-SNE visualization: The high-temperature red region identifies the paramagnetic state.The map therefore separates the disordered high-temperature configurations from the two ordered regions.
- Network action: The trained network’s low-temperature output becomes high in the cool region and crosses to a low value across the separating hyperplane.This response provides a geometric interpretation of the phase classification in the embedded representation.
Appendix C: Details of the convolutional neural network of the Ising lattice gauge
The CNN is tested to determine whether it classifies Ising lattice gauge theory states using local energetic constraints or extended topological structure. Its architecture and controlled transformations show that classification relies on local constraints, which a streamlined toy CNN can explicitly detect.
- CNN architecture: The CNN receives two-dimensional Ising spin configurations and uses convolutional filters followed by fully connected and softmax layers.The input contains N = 16 × 16 × 2 spins with σ_i = ±1; the convolutional filters use unit stride and periodic boundary conditions.
- Diagnostic test sets: The test sets modify local constraints or extended closed-loop structure to identify which features determine classification.One transformation flips spins every m = 2 or m = 8 plaquettes, while another reshuffles four pieces to preserve most local constraints but disrupt non-local loops.
- Local-constraint transformations: 50% overall test accuracy for transformed T = 0 states shows that violating the original local constraints makes the CNN classify them mostly as high-temperature states.The reported accuracies are 50% and 55% for m = 2 and m = 8, respectively.
- Loop-structure transformations: The CNN recognizes configurations with disrupted extended loops as ground states with high confidence, indicating that it does not rely on the non-local loop structure.These configurations preserve local constraints except along the stitching defects.
C A B
The convolutional neural network is analyzed through local plaquette filters and controlled modifications of ground-state configurations. Its learned and analytically specified outputs classify Ising gauge-theory states perfectly while reflecting local energetic constraints and toric-code structure.
- CNN construction: The CNN processes each plaquette with filters that determine whether its energetic constraints are satisfied.The Ising gauge theory has 16 spin configurations per plaquette, with 8 satisfying the Hamiltonian constraints.
- CNN construction: The convolutional layer uses 16 L × L filter arrays and reshapes their outputs into a 16 × L^2 vector for the fully connected output layer.The first and last groups each contain 8L^2 entries, corresponding to filters f = 1 through 8 and f = 9 through 16.
- Analytical toy model: The toy model assigns O0 = 0 and O∞ = 1 when an unsatisfied plaquette occurs, while fully satisfied configurations give O0 = 1 and O∞ = 0.These output assignments enable classification of zero- and high-temperature states.
- Classification results: 100% accuracy is obtained on the test sets both for the analytically specified classification scheme and after numerical training with ReLU units and a softmax output layer.The reported classification error is expected to be exponentially small in system volume because ground states are exponentially rare at infinite temperature.
- Toric-code connection: The zero-temperature neuron O0 behaves like the amplitudes of a toric-code ground state in the σz basis.The state described by O0 is a linear combination of all four ground states with a defined parity on the torus.
- Toric-code connection: The analysis suggests that convolutional neural networks can represent ground states with nonlocal toric-code structure despite filters individually processing local plaquettes.The construction is compared with projected entangled pair states, where local tensors project out odd-parity plaquettes.