Report on 1901.11042v2

Jeremy Goetz, Yi Zhang, Michael J. LawlerOpenalex Unknown 2020 · DOI · arXiv · OpenAlex

Abstract

Detecting the subtle yet phase defining features in Scanning Tunneling Microscopy and Spectroscopy data remains an important challenge in quantum materials. We meet the challenge of detecting nematic order from the local density of states data with supervised machine learning and artificial neural networks for the difficult scenario without sharp features such as visible lattice Bragg peaks or Friedel oscillation signatures in the Fourier transform spectrum. We train the artificial neural networks to classify simulated data of symmetric and nematic two-dimensional metals in the presence of disorder. The supervised machine learning succeeds only with at least one hidden layer in the ANN architecture, demonstrating it is a higher level of complexity than a nematic order detected from Bragg peaks, which requires just two neurons. We apply the finalized ANN to experimental STM data on CaFe2As2, and it predicts nematic symmetry breaking with dominating confidence, in agreement with previous analysis. Our results suggest ANNs could be a useful tool for the detection of nematic order in STM data and a variety of other forms of symmetry breaking.

Introduction

In quantum materials, identifying phase transitions often requires detecting subtle symmetry breaking. Nematic order—where electrons break rotational symmetry without breaking translational symmetry—is particularly difficult to observe when data is obscured by noise and disorder. Traditionally, physicists rely on Fourier-transformed Scanning Tunneling Microscopy (STM) data to find sharp signatures like Bragg peaks. However, in highly disordered systems, these features are often washed out. This report examines the work of Goetz et al., who utilized Artificial Neural Networks (ANNs) to identify nematicity in "messy" Local Density of States (LDOS) maps where traditional methods fail.

Methods and Architecture

The researchers employed supervised machine learning, training ANNs on thousands of simulated LDOS images generated from a tight-binding model. These images represented two classes: symmetric ($C_4$ symmetry) and nematic ($C_2$ symmetry), with added random impurities to simulate real-world experimental conditions.

A critical finding of the study was the necessity of non-linear processing. The team compared two architectures: a simple linear classifier with no hidden layers and a single-hidden-layer ANN containing 31 sigmoid neurons.

Figure 1: Comparison of ANN architectures: (a) a simple linear classifier with no hidden layer and (b) a more complex architecture with a single hidden layer.

The linear model (Figure 1a) failed to distinguish the phases, performing near the level of random chance. In contrast, the hidden-layer architecture (Figure 1b) successfully captured the higher-order correlations necessary to identify nematicity amidst disorder.

Key Results

The single-hidden-layer ANN achieved over 95% accuracy on simulated test sets. The model's robustness was further tested by applying it to real experimental STM data from the iron-based superconductor $CaFe_2As_2$.

Figure 6: Application of the trained ANN to experimental STM data of CaFe2As2. The histograms show the distribution of nematic confidence outputs across different coarse-graining levels.

As shown in Figure 6, the ANN predicted nematic symmetry breaking with high confidence. Even as the data was coarse-grained (downsampled), the model maintained a strong preference for the nematic classification, validating its ability to extract phase-defining features from noisy, real-world signals.

Discussion

The success of the ANN highlights a shift from "Fermi-surface sensitive" methods, which rely on specific wave vectors, to real-space correlation analysis. By automatically extracting relevant features while ignoring microscopic "junk," the ANN acts as a form of renormalization group flow. This approach is not only more robust against instrumental noise and tip-induced artifacts but also suggests a broader utility for ANNs in detecting various forms of complex symmetry breaking in condensed matter physics.

References

Goetz, J. B., Zhang, Y., & Lawler, M. J. (2020). Detecting nematic order in STM/STS data with artificial intelligence. arXiv:1901.11042v2 [cond-mat.dis-nn].

Summary updated

Abstract

Detecting the subtle yet phase defining features in Scanning Tunneling Microscopy and Spectroscopy data remains an important challenge in quantum materials. We meet the challenge of detecting nematic order from the local density of states data with supervised machine learning and artificial neural networks for the difficult scenario without sharp features such as visible lattice Bragg peaks or Friedel oscillation signatures in the Fourier transform spectrum. We train the artificial neural networks to classify simulated data of symmetric and nematic two-dimensional metals in the presence of disorder. The supervised machine learning succeeds only with at least one hidden layer in the ANN architecture, demonstrating it is a higher level of complexity than a nematic order detected from Bragg peaks, which requires just two neurons. We apply the finalized ANN to experimental STM data on CaFe2As2, and it predicts nematic symmetry breaking with dominating confidence, in agreement with previous analysis. Our results suggest ANNs could be a useful tool for the detection of nematic order in STM data and a variety of other forms of symmetry breaking.
BibTeX
@inproceedings{goetz2020report,
  author = {Jeremy Goetz and Yi Zhang and Michael J. Lawler},
  title = {{Report on 1901.11042v2}},
  booktitle = {Openalex Unknown},
  year = {2020},
  doi = {10.21468/scipost.report.1514},
}
Figure