Rapid analysis of point-contact Andreev reflection spectra via machine learning with adaptive data augmentation

Dongik Lee, Valentin G. Stanev, Xiaohang Zhang, Mijeong Kang, Ichiro Takeuchi, Seunghun LeearXiv 2025 · DOI · arXiv

Abstract

Identifying the superconducting order parameter—specifically the pairing symmetry and gap magnitude ($\Delta$)—is fundamental to understanding novel superconductors. Point-Contact Andreev Reflection (PCAR) spectroscopy is a primary tool for this characterization, yet traditional analysis relies on computationally expensive and human-intensive fitting to the Blonder-Tinkham-Klapwijk (BTK) theory. We present a machine learning approach using a 1D Convolutional Neural Network (CNN) that accelerates this analysis to sub-100 ms timescales. By employing a novel "adaptive data augmentation" strategy, the model learns to distinguish critical physical features from common experimental artifacts such as thermal noise and non-ideal conductance dips. Our results demonstrate high accuracy ($R^2 > 99\%$) across various superconducting symmetries, providing a robust path for high-throughput characterization in superconductivity and spintronics research.

Introduction

PCAR spectroscopy measures the differential conductance ($dI/dV$) at a normal metal/superconductor interface. While the technique is experimentally accessible, the resulting spectra are notoriously difficult to interpret. Standard analysis involves multi-parameter fitting to the BTK model, accounting for temperature ($T$), gap size ($\Delta$), interfacial barrier strength ($Z$), and broadening ($\Gamma$).

Real-world data often deviates from ideal theoretical curves due to thermal noise and "conductance dips" caused by non-ideal contact geometry or local heating. Traditional automated fitting algorithms frequently fail by attempting to fit these non-physical artifacts, requiring human experts to manually intervene. This bottleneck limits the speed of material discovery and characterization.

Methodology

The researchers developed a 1D CNN architecture implemented in TensorFlow and Keras. The defining feature of this work is the Adaptive Data Augmentation used during the training phase to encode human-like physical intuition into the network.

Overview of the machine learning workflow for PCAR spectra analysis. The process involves generating theoretical spectra using BTK theory, applying adaptive data augmentation (noise and dips), and training a CNN to predict physical parameters ($\Delta, Z, \Gamma$).

Synthetic Data Generation and Augmentation

The training set was built using thousands of theoretical spectra generated via BTK theory for $s$-wave, chiral $p$-wave, and $d$-wave symmetries. To make the model robust, two specific augmentation strategies were applied:

  1. Selective Noise: For unconventional symmetries, higher noise levels were added to the regions outside the superconducting gap ($|V| > \Delta$). This forces the CNN to prioritize the subtle, low-noise features within the gap that define the pairing symmetry.
  2. Artifact Simulation: For $s$-wave models, Gaussian "dips" were intentionally subtracted from the synthetic data. This teaches the network to treat such features as noise to be ignored rather than physical data to be fitted.

Results and Discussion

The trained models were validated against experimental data from known superconductors, including $YB_6$, $MgCNi_3$ ($s$-wave), $Bi/Ni$ bilayers ($p$-wave), and $PuCoGa_5$ ($d$-wave).

Experimental PCAR spectrum of $MgCNi_3$ fitted by the trained CNN. The model successfully ignores the large central dip, a common experimental artifact, to extract accurate physical parameters.

Performance and Robustness

The CNN achieved $R^2$ scores exceeding $99\%$ for $s$-wave spectra and $93\text{--}94\%$ for more complex unconventional symmetries. Most significantly, the model provides parameter predictions in less than 100 ms per spectrum. This is a massive improvement over traditional methods, which can take minutes for simple models and up to 10 minutes for unconventional symmetries requiring double integration.

As shown in the $MgCNi_3$ analysis, the model successfully "sees through" significant conductance dips that would typically derail standard least-squares fitting. The CNN correctly identifies the underlying superconducting gap despite the presence of large non-physical artifacts.

Conclusion and Future Outlook

This machine learning approach transforms PCAR analysis from a slow, manual process into a rapid, automated diagnostic. While currently focused on single-gap superconductors, future iterations could expand to multi-gap systems like $MgB_2$ or incorporate spin-polarization parameters for spintronics applications. By automating the extraction of physical parameters from "dirty" experimental data, this method paves the way for real-time analysis during experimental data acquisition.

Summary updated

Abstract

Delineating the superconducting order parameters is a pivotal task in investigating superconductivity for probing pairing mechanisms, as well as their symmetry and topology. Point-contact Andreev reflection (PCAR) measurement is a simple yet powerful tool for identifying the order parameters. The PCAR spectra exhibit significant variations depending on the type of the order parameter in a superconductor, including its magnitude ($\mathitΔ$), as well as temperature, interfacial quality, Fermi velocity mismatch, and other factors. The information on the order parameter can be obtained by finding the combination of these parameters, generating a theoretical spectrum that fits a measured experimental spectrum. However, due to the complexity of the spectra and the high dimensionality of parameters, extracting the fitting parameters is often time-consuming and labor-intensive. In this study, we employ a convolutional neural network (CNN) algorithm to create models for rapid and automated analysis of PCAR spectra of various superconductors with different pairing symmetries (conventional $s$-wave, chiral $p_x+ip_y$-wave, and $d_{x^2-y^2}$-wave). The training datasets are generated based on the Blonder-Tinkham-Klapwijk (BTK) theory and further modified and augmented by selectively incorporating noise and peaks according to the bias voltages. This approach not only replicates the experimental spectra but also brings the model's attention to important features within the spectra. The optimized models provide fitting parameters for experimentally measured spectra in less than 100 ms per spectrum. Our approaches and findings pave the way for rapid and automated spectral analysis which will help accelerate research on superconductors with complex order parameters.
BibTeX
@misc{lee2025rapid,
  author = {Dongik Lee and Valentin G. Stanev and Xiaohang Zhang and Mijeong Kang and Ichiro Takeuchi and Seunghun Lee},
  title = {{Rapid analysis of point-contact Andreev reflection spectra via machine learning with adaptive data augmentation}},
  year = {2025},
  eprint = {2503.10040},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.10040},
}
Figure