Source-linked AI summary
Machine learning vortices at the Kosterlitz-Thouless transition
Matthew J. S. Beach, Anna Golubeva, Roger G. Melko
TL;DR
The paper asks whether neural networks can classify the KT transition in the two-dimensional XY model by learning topological defects rather than bulk features. It compares raw-spin, vortex, and custom-network approaches, finding that finite-size magnetization can support raw-input classification while learning vortices remains difficult, especially at smaller sizes.
Problem
The paper investigates whether neural networks can distinguish phases of the XY model using topological defects, and whether raw spin states can yield features interpretable as local vorticity.
Source: §I · p. 1, §III · p. 2Method
The authors compare supervised networks on raw spins and preprocessed vorticity, use a confusion scheme, and design a deep architecture to learn vortices from raw configurations.
Source: §I · p. 1, B. Learning by confusion · p. 4, §IV · p. 7Results
Raw-input convolutional networks classify consistently well as lattice size increases, with evidence pointing to finite-size magnetization; vortex preprocessing yields finite-size scaling, but the one-layer network plateaus near L = 32 and vortex confusion peaks near T* ≈ 1.
Source: §IV · p. 7Takeaways & Limitations
Feature engineering and finite lattice size materially affect supervised learning of the KT transition, and vortex detection does not necessarily produce the best classification accuracy.
Source: §IV · p. 7For systems above 1,000 spins, vortex learning remains difficult: randomly initialized networks can settle in local minima unrelated to vortices, and the designed optimization landscape may make vortex minima hard to find.
Source: §IV · p. 7, C. Custom architecture for learning… · p. 7
Abstract
from arXiv · showhide
Efficient and automated classification of phases from minimally processed data is one goal of machine learning in condensed matter and statistical physics. Supervised algorithms trained on raw samples of microstates can successfully detect conventional phase transitions via learning a bulk feature such as an order parameter. In this paper, we investigate whether neural networks can learn to classify phases based on topological defects. We address this question on the two-dimensional classical XY model which exhibits a Kosterlitz-Thouless transition. We find significant feature engineering of the raw spin states is required to convincingly claim that features of the vortex configurations are responsible for learning the transition temperature. We further show a single-layer network does not correctly classify the phases of the XY model, while a convolutional network easily performs classification by learning the global magnetization. Finally, we design a deep network capable of learning vortices without feature engineering. We demonstrate the detection of vortices does not necessarily result in the best classification accuracy, especially for lattices of less than approximately 1000 spins. For larger systems, it remains a difficult task to learn vortices.
I. INTRODUCTION
The paper asks whether supervised neural networks can identify the two-dimensional XY model’s KT transition from topological defects, rather than relying on bulk features. It compares existing learning approaches and tests raw and processed inputs and architectures intended to learn vortices.
I. INTRODUCTION
- The motivation is to test whether neural networks can distinguish unconventional transitions driven by topological defects, extending beyond symmetry-breaking cases where learning an order parameter can explain classification.Earlier classical statistical-physics studies found that networks could learn order parameters or other thermodynamic quantities [6] [8].Source: §I · p. 1
- Unsupervised methods applied to the XY model often learn finite-size magnetization; PCA applied to vorticity does not resolve vortex–antivortex unbinding.The cited work attributes PCA’s difficulty to its linearity, while variational auto-encoders classify by learning bulk magnetization.Source: §I · p. 1
- Prior supervised studies found that a convolutional network trained on winding numbers classified boson phases across a KT transition, but the same method failed on raw configurations [16].A separate study inferred that a network trained on topological-band-insulator Hamiltonians learned to calculate global winding number [17].Source: §I · p. 1
- The study tests whether networks trained on phase-labeled XY spin states can learn a representation interpretable as local vorticity and classify the KT transition.It compares feed-forward and convolutional networks on raw spins and vorticity, and explores confusion learning and deeper architectures.Source: §I · p. 1
- Reasonable classification accuracy can arise from a local optimization minimum unrelated to topological features, so accuracy alone need not demonstrate vortex learning.The paper also frames detecting vortex–antivortex unbinding with machine learning as challenging.Source: §I · p. 2
II. BACKGROUND
In the two-dimensional XY model, vortex–antivortex unbinding separates a low-temperature quasi-long-range-ordered phase from a high-temperature disordered phase. The section defines the defects and describes finite-size estimates of the transition temperature.
II. BACKGROUND
- Vorticity measures the winding of spin angles around a closed path; on a lattice, angle differences around a plaquette approximate the circulation, with k = 1 for a vortex and k = −1 for an antivortex.Figure 1 illustrates a vortex–antivortex pair.Source: §II · p. 2, Figure 1 · p. 2
- At TKT, vortex–antivortex pairs unbind, separating the low-temperature quasi-LRO phase with algebraically decaying correlations from the high-temperature disordered phase with exponential decay.Below TKT, thermally formed pairs remain bound; the transition occurs when entropy balances their binding energy.Source: §II · p. 2
- The finite-size critical temperature is estimated from the Nelson–Kosterlitz universal jump, where helicity modulus Υ crosses 2T/π, and scales toward the thermodynamic limit as (log L)^-2.The helicity modulus measures the response to a twist in boundary conditions.Source: §II · p. 2
- Finite-size systems have nonzero magnetization at T > 0, and at the critical point it scales as L^-1/8 to within 4% error.The simulations estimate TKT = 0.899±0.06, consistent with the literature value 0.893.Source: §II · p. 2
III. METHODS & RESULTS
The methods test binary classification of XY configurations labeled below or above TKT, with attention to whether networks use finite-size magnetization or topological defects. Training compares feed-forward and convolutional architectures on Monte Carlo data.
III. METHODS & RESULTS
- The classification task assigns configurations to low- or high-temperature phases and asks whether networks rely on finite-size magnetization or topological defects.The study also asks which architectures and features support this classification.Source: §III · pp. 2–3
- Training data comprise 1,000 configurations at each of 64 temperatures from 0.1 to 3.0 for lattice sizes L = 8 through 64; separate tests use 100 configurations per temperature.Ten percent of training data are held out for cross-validation and early stopping.Source: §III · p. 3
- Weights and biases are optimized by back-propagation against cross-entropy loss, with training stopped when validation loss fails to decrease for 50 epochs.The loss compares predicted and true binary labels.Source: §III · p. 3
- The feed-forward network has one hidden layer of 1,024 sigmoid units, while the CNN uses two convolutional layers followed by max-pooling and dense classification.The CNN exploits the input’s two-dimensional structure; the passage reports 1,024L^2 + 2,057 trainable parameters for the FFNN and 128L^2 − 1,024L + 3,361 for the CNN.Source: §III · p. 3
A. Finite-size scaling of supervised learning
The study compares phase classification and transition-temperature estimates from raw XY spin configurations with estimates from engineered vorticity inputs. Raw-spin CNN accuracy improves with system size but shows no clear finite-size scaling, while vorticity-trained networks scale toward the correct TKT.
A. Finite-size scaling of supervised learning
- The predicted critical temperature is the temperature where the sigmoid output crosses 0.5, and variance estimates use ten training runs with different validation sets.The figure caption specifies early stopping after validation loss fails to improve for 50 epochs.Source: A. Finite-size scaling of supervised… · p. 3, Figure 3 · p. 4
- For raw spins, the FFNN has low accuracy for L > 48 and a poor TKT estimate, while CNN accuracy improves without a clear finite-size scaling trend.The authors suggest the CNN may use a robust global feature rather than topological information.Source: A. Finite-size scaling of supervised… · p. 4, Figure 3 · p. 4
- The authors argue finite-size magnetization can let a CNN classify raw spins, although global magnetization is irrelevant to TKT in the thermodynamic limit.They conclude that feature engineering is crucial for claiming that classification relies on topological defects.Source: A. Finite-size scaling of supervised… · p. 4
- Vorticity inputs are constructed by wrapping neighboring-spin angle differences into (−π, π] and summing the rescaled differences around plaquettes.Raw inputs are lattice angle values with periodic boundary conditions.Source: A. Finite-size scaling of supervised… · pp. 3–4
- For vorticity inputs, both networks achieve high accuracy and scale toward the correct TKT, although the FFNN performs poorly for L > 32 while the CNN continues improving.The scaling alone does not identify whether the CNN uses squared vorticity or a more complex feature such as vortex–antivortex separation.Source: A. Finite-size scaling of supervised… · p. 4, Figure 3 · p. 4
B. Learning by confusion
The learning-by-confusion scheme trains classifiers on deliberately varied phase boundaries and evaluates their test accuracy. Raw-spin and vorticity inputs produce different accuracy patterns, but finite-size effects and stronger performance at a false boundary complicate transition identification.
B. Learning by confusion
- Learning by confusion varies a proposed boundary T* to create supervised labels, trains a network for each labeling, and plots test accuracy against T*.The scheme assumes the network can learn the true physical labeling more accurately than false labelings; endpoint peaks arise trivially when training and testing use one class.Source: B. Learning by confusion · pp. 4–5
- With raw spins, the CNN produces a −∨ rather than the expected ∨∨ accuracy shape, consistent with classification by a finite-size magnetization threshold.This supports the authors’ earlier hypothesis that raw-spin training leads the CNN to learn magnetization.Source: B. Learning by confusion · p. 5
- With vorticity inputs, the expected ∨∨ shape appears but is skewed; the network performs best near TKT, where vortex unbinding separates the phases.Low-temperature configurations have few, logarithmically bound vortices, while high-T* configurations are sufficiently random to be misclassified for arbitrary partitions.Source: B. Learning by confusion · p. 5
- For L = 64, confusion accuracy is higher near the false boundary T* ≈ 1 than at TKT = 0.89, while raw-spin accuracy exceeds 98.5% for every T* < TKT.The authors note significant finite-size effects and that the false-boundary effect might disappear in the thermodynamic limit.Source: B. Learning by confusion · p. 5
- The authors report no strong evidence that their CNN can find vortices, despite slightly higher near-TKT accuracy for the vortex-trained network.They propose a custom network designed for vortex detection as a follow-up test.Source: B. Learning by confusion · p. 5
C. Custom architecture for learning vortices
The custom architecture tests whether an intermediate vorticity representation can be learned for XY phase classification. It helps only beyond L>32; smaller lattices perform worse with fixed vortex features.
C. Custom architecture for learning vortices
- A three-layer custom network targets vorticity as an intermediate representation, using four 2×2 filters for plaquette angle differences, sawtooth wrapping, then a 1×1 sum.The sum of the four shifted, wrapped angle differences computes vorticity.Source: C. Custom architecture for learning… · pp. 5–6, Figure 5 · p. 6
- The authors compare fixed, hand-engineered vorticity weights; vortex-initialized weights that are relaxed during training; and random initialization.The fixed network provides a benchmark, while the initialized variant tests whether the vortex solution is stable.Source: C. Custom architecture for learning… · p. 6
- For L<16, fixed-vorticity weights perform poorly, but for L>32 the fixed network outperforms the other variants, making vortex representations beneficial only at larger sizes.The test-loss comparison identifies an approximate crossover near L∼32.Source: Figure 6 · p. 6, C. Custom architecture for learning… · p. 6
- All three variants are trained for binary classification with cross-entropy, using ten runs with different validation sets and early stopping after 50 unimproved epochs.Training spans lattice sizes L=8 through 72 in increments of 8.Source: C. Custom architecture for learning… · p. 6
- Histogram checks show the vortex-initialized network learns quantized vorticity only for L≥32, while random initialization yields no vortex-consistent distribution at any studied size.For small L, the initialized network's vorticity output is not quantized.Source: C. Custom architecture for learning… · pp. 6–7, Figure 7 · p. 7
- For large L, the vortex solution is a stable local minimum, but global optimality is unresolved; integral-output regularization peaked near zero and nearest-neighbor constraints were too restrictive.The authors suggest regularization might alter the optimization landscape, but report unsuccessful attempts with these approaches.Source: C. Custom architecture for learning… · p. 7
IV. CONCLUSION
Learning the XY model’s KT transition depends strongly on input representation and lattice size: raw-spin success can reflect magnetization, while vortex learning remains difficult.
IV. CONCLUSION
- The results demonstrate that feature engineering and finite lattice sizes strongly affect supervised-learning performance on the KT transition.The authors argue that the difficulty of supervised phase discrimination also underscores challenges for unsupervised learning from unlabeled data.Source: §IV · p. 7
- On raw spins, the feed-forward network failed to estimate TKT accurately for moderate lattices, while the convolutional network performed consistently well and was inferred to learn finite-size magnetization.Raw-spin training in the confusion method also readily learned false phase boundaries below TKT.Source: §IV · p. 7
- With vorticity preprocessing, both architectures showed finite-size scaling consistent with thermodynamic TKT; the convolutional network improved with size, while the one-layer network plateaued near L = 32.The vortex-trained confusion scheme instead peaked near T* ≈ 1 rather than predicting the correct critical temperature.Source: §IV · p. 7
- Vortex discovery benefited the deep network only above 1000 spins, and random initialization still settled into non-vortex local minima.The authors suggest the optimization landscape may be sufficiently rough that stochastic gradient descent takes exponentially long to find vortex-related minima.Source: §IV · p. 7