Deciphering Majorana Zero Modes in Topological Superconductor FeTe0.55Se0.45 with Machine-Learning-Assisted Spectral Deconvolution

arXiv:2602.15178 · cond-mat.supr-con · Submitted 2026-02-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: Today's paper: "Deciphering Majorana Zero Modes in Topological Superconductor FeTe0.55Se0.45 with Machine-Learning-Assisted Spectral Deconvolution".

Mira: Unambiguous identification of Majorana zero modes (MZMs) in topological superconductors (TSCs) remains a challenge due to complex in-gap states that can also produce zero-bias conductance peaks (ZBPs).

Kai: First, who's behind it and why it matters.

Title and authors: Kai: So we're looking at the paper titled "Deciphering Majorana Zero Modes in Topological Superconductor FeTe0 point 55Se0 point 45 with Machine-Learning-Assisted Spectral Deconvolution <ref:2602.15178#pg0>." It sounds like they've tackled a really messy problem in identifying genuine Majorana zero modes, which is tough because there are so many other states that can mimic those zero-bias peaks.

Mira: I think the title immediately signals that this paper isn't just looking at any superconductor; it’s specifically focusing on an intrinsic topological superconductor called FeTe0 point 55Se0 point 45 and using a machine learning approach to separate the real Majorana signatures from those tricky trivial states <ref:2602.15178#pg0>.

Lev: From an error correction standpoint, isolating the true zero-energy modes is critical because if we can't tell them apart, we can't reliably map out the system for any kind of topological qubit implementation one <ref:2602.15178#pg0>.

Kai: Exactly. It’s about moving past just seeing a peak and actually understanding what kind of peak it is before we try to build anything on top of it.

Mira: And the authors are setting up this workflow to take complex local density of states data, decompose it into individual spectral components, and then use machine learning to classify those components based on their properties.

Lev: That classification step is where the real hardware relevance comes in; if the ML model can reliably distinguish between a true zero-energy mode and a trivial excitation, then we have something that could actually be engineered for error suppression one <ref:2602.15178#pg0>.

The paper's summary: Kai: To summarize what this paper is doing, they are taking tunneling spectroscopy data from the FeTe0 point 55Se0 point 45 TSC and applying a data-driven workflow that combines pixel-wise spectral deconvolution with machine learning to objectively separate actual Majorana zero modes from those trivial in-gap states that can also cause zero-bias conductance peaks <ref:2602.15178#pg0>.

Mira: It’s essentially taking the raw, noisy LDOS data and breaking it down into individual Lorentzian components, extracting parameters for each peak—like its center and width—and then feeding those features into an unsupervised machine learning model to sort them into categories.

Lev: That decomposition step is key because the paper mentions that simple inspection isn't enough; they are using this structured feature set F to feed the ML, which suggests a systematic way to handle the complexity of multi-peak spectra <ref:2602.15178#pg1>.

Kai: Right. They are not just looking at one spectrum; they're analyzing a grid of spectra across the material, trying to find patterns that point toward MZM signatures rather than just random noise or trivial excitations.

Mira: The paper highlights that this method helps rule out several non-topological mechanisms, which means they are specifically focusing on strengthening the ZBP as an MZM signature by explicitly excluding other possibilities <ref:2602.15178#pg2>.

Lev: If the ML can successfully classify these features, then it provides a way to filter the experimental noise and focus only on the signals that matter for topological physics, which is exactly what we need for running any kind of fault-tolerant computation one <ref:2602.15178#pg0>.

The paper's improvements: Kai: One of the main improvements they are pushing is moving beyond just looking at isolated zero-bias peaks and instead developing a workflow that integrates spatial resolution with spectral deconvolution to analyze tunneling spectroscopy from FeTe0 point 55Se0 point 45 in a data-driven way <ref:2602.15178#pg1>.

Mira: The core improvement lies in the methodology: they use pixel-wise spectral deconvolution, where each local dI/dV spectrum is decomposed into multiple Lorentzian peaks, and then those extracted parameters are fed into UMAP embedding followed by HDBSCAN clustering to classify them <ref:2602.15178#pg1>.

Lev: The paper points out that they augment this feature set F with features that capture "zero-bias proximity and spectral symmetry," which is a smart way to ensure the ML isn't just looking at generic peak shapes but is actually looking for the topological physics <ref:2602.15178#pg1>.

Kai: So, it’s not just about fitting peaks; it’s about engineering a feature set that specifically highlights what we think is important for identifying MZMs in this specific material system, FeTe0 point 55Se0 point 45 <ref:2602.15178#pg1>.

Mira: And the result of this classification is that they identify a cluster C0 which is "sharply concentrated near zero-energy" and localized around vortex cores, while other clusters, C1 and C2, show broader distributions in space and energy <ref:2602.15178#pg4>.

Lev: That spatial differentiation between the clusters is significant because it suggests that not all observed zero-bias features are the same; some are localized vortex cores, while others are more diffuse states, which gives us a concrete physical distinction <ref:2602.15178#pg4>.

Conclusion: Kai: So to wrap up on this paper, they’ve developed a reproducible and scalable framework that uses machine learning to reliably extract MZM spectral features from complex LDOS data, which successfully mitigates misidentification arising from trivial near-zero-energy states <ref:2602.15178#pg4>.

Mira: The conclusion is that this method allows them to reconstruct a ZBP-only local density of states map, rho ZBP(E, r), by summing only the components assigned to Cluster C0, which shows that only a subset of vortices exhibit those pronounced circular-shaped zero-bias enhancements consistent with established experimental signatures of MZMs <ref:2602.15178#pg4>.

Lev: For error correction researchers, the implication is that this provides a clear path for mapping out the actual topological regions in hardware, allowing us to focus our efforts on where those robust ZBPs are located and where we need to worry about suppression due to disorder <ref:2602.15178#pg4>.

Kai: It’s a big step in making this analysis scalable, moving it from analyzing a limited subset of data to classifying the entire grid LDOS dataset using this ML approach <ref:2602.15178#pg4>.

Mira: This framework opens the door for applying these ideas to other topological heterostructure systems and even allows for classification across diverse materials, which is a significant extension of this specific work <ref:2602.15178#pg4>.

Lev: If we can establish this kind of automated classification, then we gain a tool that could potentially be used on real-time experimental data streams to rapidly assess the topological fidelity of new material interfaces one <ref:2602.15178#pg0>.

Center for Nanophase Materials Sciences, Oak Ridge National Laboratory

cond-mat.supr-con

Submitted: 2026-02-16

Updated: 2026-05-30

DOI: 10.1038/s42005-026-02828-9

Code: https://github.com/jewook-park/ZBPs_in_FTS

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: Unambiguous identification of Majorana zero modes (MZMs) in topological superconductors (TSCs) remains a challenge due to complex in-gap states that can also produce zero-bias conductance peaks

Key concepts

Majorana Zero Modes (MZMs)
These are elusive zero-energy electronic states predicted in topological superconductors. Identifying them is difficult because they can be confused with other trivial states that also show zero-bias conductance peaks, requiring sophisticated analysis to confirm their true topological nature.
Spectral Deconvolution
This technique breaks down a complex measured signal (the local dI/dV spectrum) into its constituent simple components, usually modeled as multiple Lorentzian peaks. This allows researchers to separate overlapping spectral features that arise from different physical states within the material.
Unsupervised Machine Learning
The study used algorithms like HDBSCAN and UMAP to automatically classify the extracted spectral features without prior labels. The ML system learned patterns in the data, allowing it to group peaks based on their spatial and energy characteristics, helping separate MZM signatures from noise.

Terminology

Summary

Unambiguous identification of Majorana zero modes (MZMs) in topological superconductors (TSCs) remains a challenge due to complex in-gap states that can also produce zero-bias conductance peaks (ZBPs). This study demonstrates a data-driven workflow integrating pixel-wise spectral deconvolution with machine learning to analyze tunneling spectroscopy from FeTe0.55Se0.45, an intrinsic TSC, to objectively separate genuine MZM signatures from trivial in-gap states.

The gist

A data-driven workflow combining high spatial and energy resolution STM/S data acquisition, spectral deconvolution with multiple peak fitting, and unsupervised machine learning is used to isolate ZBPs consistent with Majorana zero modes (MZMs) in FeTe0.55Se0.45 by classifying features extracted from grid local density of states (LDOS) data.

Experimental Setup and Data Acquisition

The research utilized a dilution refrigerator STM operating at 40 mK equipped with a vector field magnet (2-2-9 T). The FeTe0.55Se0.45 single crystal was cleaved at 83 K under ultra-high vacuum and immediately transferred to a precooled STM head (4.2 K), reaching the base temperature of 40 mK after further cooling with the dilution refrigerator. High energy resolution was achieved to distinguish expected CdGM levels and ZBPs. The study focused on atomically clean surfaces without Fe adatoms or domain boundaries to avoid contributions from local heterogeneities, confirmed by STM topography and fast Fourier transform analysis showing the absence of excess Fe adatoms and domain boundaries.

Spectral Deconvolution and Feature Engineering

The grid LDOS data, acquired as a function of energy and spatial coordinates, was analyzed to extract spatially resolved spectral features. Each local dI/dV spectrum was decomposed into a sum of multiple Lorentzian peaks:

  1. The extracted peak parameters—center (cijk), amplitude (aijk), and width (wijk)—were assembled into a structured feature set, denoted as F.

  2. This feature set was augmented with additional features that capture zero-bias proximity and spectral symmetry.

  3. These parameters were then processed through UMAP embedding followed by hierarchical density-based spatial clustering of applications with noise (HDBSCAN).

Machine Learning Classification

Unsupervised ML algorithms were employed to classify the deconvoluted peak components. The process involved:

  1. Statistical outliers were excluded using Principal Component Analysis (PCA) and k-nearest neighbor (kNN) distance scoring.

  2. Cleaned features were embedded using Uniform Manifold Approximation and Projection (UMAP).

  3. Clustering was performed using HDBSCAN, resulting in three data-driven clusters: C0, C1, and C2.

  4. A cluster consistent with ZBPs was identified through quantitative cluster statistics, leading to the assignment of corresponding deconvoluted peaks as ZBP components.

Identification and Validation of MZM Signatures

The classification successfully distinguished ZBP-consistent features from trivial in-gap states:

  1. The analysis revealed that Cluster C0 is sharply concentrated near zero-energy and localized around vortex cores, while C1 and C2 exhibit broader distributions in space and energy.

  2. Reconstructing the ZBP-only LDOS, denoted as rhoZBP(E, r), by summing the Lorentzian components assigned to Cluster C0 effectively removes non-ZBP components.

  3. The resulting map showed that only a subset of vortices exhibited pronounced, circular-shaped ZBC enhancements (red circles in Fig. 4f), consistent with established experimental signatures of MZMs, while others were attributed to trivial in-gap states (blue circles).

  4. Furthermore, the study correlated local heterogeneity with ZBP formation; vortices with reduced ZBPs tended to be located closer to subsurface defects, suggesting an association between local heterogeneity and the suppression or distortion of vortex-core ZBPs.

Conclusion and Future Directions

The data-driven workflow provides a reproducible and scalable framework for reliably extracting MZM spectral features from complex LDOS data. This method successfully mitigates misidentification arising from trivial near-zero-energy states by classifying spectral properties. The workflow can be extended by applying deep-learning architectures or classifying complex in-gap states across diverse topological heterostructure systems, and it can be adapted to complementary probes such as nonlocal transport and spin-polarized STM to strengthen MZM identification. This provides a robust foundation for the systematic mapping and manipulation of MZMs in TSCs.

References

  1. Kitaev, A.Y., Fault-tolerant quantum computation by anyons. Annals of Physics, 2003. 303(1): p. 2-30.

Improvements for AI systems

Here are the specific improvements that can be made to AI systems, directly informed by the methodology described in this scientific paper, and what those improved AI systems could achieve:

  1. A data-driven workflow integrating pixel-wise spectral deconvolution with unsupervised Machine Learning (ML) for analyzing tunneling spectroscopy data from intrinsic Topological Superconductors (TSCs).

  2. An ML system capable of decomposing local density of states (LDOS) spectra acquired with a millikelvin Scanning Tunneling Microscope (STM) under magnetic fields into multiple Lorentzian peaks, extracting peak parameters, and assembling them into a structured feature set.

  3. An unsupervised ML clustering algorithm that identifies distinct classes of LDOS spectra based on extracted peak parameters, specifically separating superconducting vortices exhibiting Zero-Bias Peaks (ZBPs) consistent with Majorana Zero Modes (MZMs) from vortices displaying ZBP-mimicking features of trivial origin.

  4. A spatially resolved ML analysis capable of differentiating between isotropic vortex cores with well-defined ZBPs and vortices that exhibit locally distorted ZBPs by analyzing the spatial distribution of these features across the grid LDOS data.

  5. An ML-assisted workflow that can systematically organize and classify deconvoluted peak components across entire grid LDOS datasets to disentangle MZM-related ZBPs from trivial vortex core states, moving beyond limited subset analysis (line or point spectroscopy).

  6. A system capable of reconstructing a ZBP-only local density of states map, filtering out non-ZBP components using ML-assigned cluster labels (e.g., isolating Cluster C0), to visualize the spatial and energetic distribution of genuine MZM signatures while effectively removing features from trivial in-gap states (C1, C2).

  7. A system that correlates local heterogeneity (e.g., subsurface defects identified via zero-field LDOS maps) with the suppression or distortion of vortex core ZBPs, providing a predictive model for MZM robustness based on spatial proximity to disorder.

These improved AI systems can achieve the following:

  1. Predictive Identification of Topological States: The system can reliably and objectively identify true MZM signatures in experimental STM/S data, significantly reducing false positives caused by trivial near-zero-energy states (CdGM levels, YSR states).

  2. High-Resolution Spatial Mapping: It can generate high-fidelity 3D maps (LDOS) that spatially resolve the energy distribution of quasiparticles, allowing researchers to precisely locate and characterize MZM vortex cores in real space.

  3. Automated Feature Extraction Pipeline: The system automates the complex, multi-step process of spectral deconvolution and feature engineering from raw spectroscopic data, providing a scalable framework for analyzing large experimental datasets without requiring exhaustive manual inspection of every local spectrum.

  4. Quantification of Topological Robustness: By correlating spatial proximity to disorder with ZBP suppression, the AI can quantify how material heterogeneity affects the stability and observable signature of MZMs in TSCs.

  5. Scalable Discovery Framework: The resulting workflow establishes a reproducible and scalable framework applicable to diverse topological heterostructure systems, enabling rapid screening of new materials for MZM detection based on spectral signatures rather than just isolated zero-bias peaks.

Sources

Related papers