Report on 1901.11042v1

Jeremy Goetz, Yi Zhang, Michael J. LawlerOpenalex Unknown 2019 · 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 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 isotropic and anisotropic 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 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 99% confidence (probability 0.99), 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.

The Challenge of Disordered Quantum Materials

In quantum materials research, Scanning Tunneling Microscopy (STM) is used to map the Local Density of States (LDOS), providing a spatial visualization of electronic behavior. A key objective is identifying nematic order, a phase where electronic systems break the crystal's four-fold symmetry ($C_4$) to favor a two-fold symmetry ($C_2$).

However, real-world materials are often "dirty," containing impurities and noise that obscure these symmetries. Traditional methods, such as Fourier transforms, rely on sharp features like Bragg peaks. In highly disordered systems, these features vanish, leaving a "blurry mess" that standard algorithms and human intuition struggle to interpret. This research investigates whether Artificial Neural Networks (ANNs) can extract these hidden symmetries where traditional techniques fail.

Methodology: Beyond Linear Analysis

The researchers employed supervised machine learning, comparing a simple linear model against a non-linear ANN with a hidden layer of 31 sigmoid neurons.

Comparison of (a) linear and (b) non-linear ANN architectures used for classification.

To train the models, a dataset of ~5,000 simulated LDOS images was generated using tight-binding models. These simulations incorporated varying levels of "quenched disorder" (3% to 20% impurity concentration) to accurately mimic experimental conditions. The models were implemented using TensorFlow, with gradient descent used to optimize weights and biases.

Key Insights: The "Hidden" Intelligence of ANNs

A critical finding was that the linear ANN performed no better than random chance (53% accuracy). This indicates that detecting nematicity in disordered systems is a "higher-level" complexity problem requiring the non-linear correlations provided by a hidden layer.

Analysis of hidden neuron firing patterns relative to output weights, showing how the network distinguishes between phases.

By analyzing the hidden neurons, researchers found that the AI does not rely on a single visual feature. Instead, it develops a complex interplay of patterns. The authors suggest the ANN acts similarly to a renormalization group flow, automatically filtering out microscopic noise to focus on macroscopic symmetry.

Results and Experimental Validation

The non-linear model achieved 95% accuracy on simulated data, significantly outperforming Fourier analysis in high-disorder regimes. More importantly, when applied to real experimental data from the iron-based superconductor $\text{CaFe}_2\text{As}_2$, the model predicted nematic symmetry breaking with 99% confidence. This prediction held even when the data was downsampled, proving the model learned fundamental physics rather than pixel-level noise.

Future Outlook

While highly successful, the current model cannot yet distinguish between intrinsic material nematicity and instrumental artifacts, such as an asymmetric microscope tip. Future work involves training models with additional categories to help the AI learn to subtract instrumental bias from physical reality. Nevertheless, the success of this approach suggests that ANNs are powerful tools for detecting elusive symmetry breaking and topological phases in a wide range of quantum materials.

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 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 isotropic and anisotropic 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 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 99% confidence (probability 0.99), 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{goetz2019report,
  author = {Jeremy Goetz and Yi Zhang and Michael J. Lawler},
  title = {{Report on 1901.11042v1}},
  booktitle = {Openalex Unknown},
  year = {2019},
  doi = {10.21468/scipost.report.1488},
}
Figure