Abstract
We train a neural network to identify impurities in experimental images obtained by scanning tunneling microscope (STM) measurements. The neural network is first trained with a large number of simulated data and then applied to identify experimental images taken at different voltages. We use a convolutional neural network to extract features and implement an attention mechanism to capture correlations between images at different voltages. Simulated data captures universal Friedel oscillations but often fails to describe non-universal short-range physics or experimental noise. We demonstrate that including uncorrelated white noise in simulations and developing a confidence-based loss function for experimental data significantly improves performance.
Introduction
Scanning Tunneling Microscopy (STM) is a powerful tool for visualizing electron density on material surfaces at the atomic scale. Impurities, such as foreign atoms, disrupt this density, creating ripples known as Friedel oscillations. While theoretical models can describe the long-range behavior of these oscillations, the physics in the immediate vicinity of an impurity—the "short-range" physics—is often non-universal and difficult to model accurately.
Applying deep learning to identify these impurities faces a significant "reality gap." Labeled experimental data is scarce, necessitating the use of synthetic training data. However, models trained on idealized simulations often fail when applied to noisy, complex experimental images. This research addresses how to bridge this gap using specialized neural network architectures and training strategies.
Methodology
The researchers developed a framework centered on a Convolutional Neural Network (CNN) designed to extract spatial features from STM images. To handle the multi-dimensional nature of STM data, where images are captured across a range of bias voltages, an attention mechanism was integrated. This allows the network to correlate features that persist across different energy levels, effectively distinguishing static physical impurities from transient noise.
A core innovation is the introduction of Confidence Regularization. As shown in Figure 1, the training process utilizes both theoretical data (with known labels) and experimental data (without labels). For the experimental branch, a regularization loss ($\mathcal{L}_{reg}$) is calculated based on the entropy of the network's predictions. This encourages the model to produce "sharp," confident predictions on real data, preventing it from overfitting to unphysical artifacts present only in the simulations. Additionally, the researchers found that augmenting simulated data with simple Gaussian white noise forced the network to prioritize robust physical patterns over pixel-level perfection.
Datasets and Evaluation
The model was trained on a large-scale synthetic dataset generated from theoretical Friedel oscillation equations, incorporating various impurity positions and scattering properties.
For testing, real STM measurements of a Gold (Au(111)) surface were used, consisting of image sequences taken at bias voltages from -230 meV to +20 meV. To validate the model's performance, the researchers compared its predictions against a reference measurement taken at -520 meV. At this specific voltage, impurities appear as distinct "bound states," providing a reliable ground truth for evaluating the network's accuracy on the noisier, higher-voltage data.
Results and Discussion
The study compared four distinct training configurations to evaluate the impact of noise augmentation and confidence regularization.
As illustrated in Figure 4, the baseline model trained only on "raw" clean simulations performed poorly on experimental data. Adding noise augmentation significantly boosted the identification ratio. While regularization alone provided some benefit, the combined approach (noise + regularization) achieved the highest performance. This synergy demonstrates that forcing the model to ignore noise while simultaneously rewarding confident predictions on real-world data is the most effective way to bridge the simulation-to-reality gap.
Qualitatively, the final model successfully identified impurity locations that were nearly indistinguishable from noise to the human eye, with high spatial correlation to the reference bound-state data.
Conclusion
This work demonstrates a robust method for applying machine learning to experimental physics data when labeled samples are unavailable. By combining physics-informed architectures (attention) with strategies to handle the reality gap (noise augmentation and confidence regularization), the researchers successfully automated the identification of atomic-scale defects. This approach holds promise for analyzing more complex quantum materials, such as high-temperature superconductors, where defect identification is crucial for uncovering new physical phenomena.
Summary updated