Self-supervised learning for denoising quasiparticle interference data

TL;DR

This paper introduces self-supervised machine learning methods to effectively denoise quasiparticle interference data, improving clarity and analysis without requiring noiseless training examples, applicable to both simulated and real experimental data.

Contribution

It adapts Noise2Noise and Noise2Self algorithms for quasiparticle interference data, demonstrating superior denoising performance over traditional methods in both simulated and experimental contexts.

Findings

Effective noise reduction while preserving details

Enhanced clarity of interference patterns in experimental data

Facilitates more accurate electronic structure analysis

Abstract

Tunneling spectroscopy is an important tool for the study of both real-space and momentum-space electronic structure of correlated electron systems. However, such measurements often yield noisy data. Machine learning provides techniques to reduce the noise in post-processing, but traditionally requires noiseless examples which are unavailable for scientific experiments. In this work we adapt the unsupervised Noise2Noise and self-supervised Noise2Self algorithms, which allow for denoising without clean examples, to denoise quasiparticle interference data. We first apply the techniques on simulated data, and demonstrate that we are able to reduce the noise while preserving finer details, all while outperforming more traditional denoising techniques. We then apply the Noise2Self technique to experimental data from an overdoped cuprate ((Pb,Bi)$_2$Sr$_2$CuO$_{6+\delta}$) sample. Denoising enhances the clarity of quasiparticle interference patterns, and helps to obtain a precise extraction of electronic structure parameters. Self-supervised denoising is a promising tool for denoising quasiparticle interference data, facilitating deeper insights into the physics of complex materials.