Weak-signal extraction enabled by deep-neural-network denoising of diffraction data

Jens Oppliger,M. Michael Denner,Julia Küspert,Ruggero Frison,Qisi Wang,Alexander Morawietz,Oleh Ivashko,Ann-Christin Dippel,Martin von Zimmermann,Izabela Biało,Leonardo Martinelli,Benoît Fauqué,Jaewon Choi,Mirian Garcia-Fernandez,Ke-Jin Zhou,Niels B. Christensen,Tohru Kurosawa,Naoki Momono,Migaku Oda,Fabian D. Natterer,Mark H. Fischer,Titus Neupert,Johan Chang

DOI: https://doi.org/10.1038/s42256-024-00790-1

2023-12-12

Abstract:Removal or cancellation of noise has wide-spread applications for imaging and acoustics. In every-day-life applications, denoising may even include generative aspects, which are unfaithful to the ground truth. For scientific use, however, denoising must reproduce the ground truth accurately. Here, we show how data can be denoised via a deep convolutional neural network such that weak signals appear with quantitative accuracy. In particular, we study X-ray diffraction on crystalline materials. We demonstrate that weak signals stemming from charge ordering, insignificant in the noisy data, become visible and accurate in the denoised data. This success is enabled by supervised training of a deep neural network with pairs of measured low- and high-noise data. We demonstrate that using artificial noise does not yield such quantitatively accurate results. Our approach thus illustrates a practical strategy for noise filtering that can be applied to challenging acquisition problems.

Image and Video Processing,Strongly Correlated Electrons,Superconductivity,Machine Learning

What problem does this paper attempt to address?

This paper attempts to address the problem of how to remove noise from low-count (LC) data in X-ray diffraction (XRD) experiments using deep neural networks, specifically convolutional neural networks (CNN), to extract weak signals. Specifically, the authors focus on charge density wave (CDW) signals that are difficult to discern against a high-noise background.

### Background and Challenges
- **Importance of Noise Removal**: Noise removal has widespread applications in imaging and acoustics. For everyday applications, denoising may include generative aspects that are not entirely faithful to the real situation. However, in scientific research, denoising must faithfully reproduce the real situation.
- **Characteristics of X-ray Diffraction Data**: XRD data typically contain multiple noise sources, such as Poisson noise and readout noise, which have different statistical properties. Therefore, it is challenging to accurately simulate experimental noise.
- **Challenges of Low-Count Data**: Under certain experimental conditions, long exposure times are not possible due to radiation damage, limited counting time under pulsed magnetic fields, or insufficient statistics when exploring multi-dimensional parameter spaces. This leads to the generation of low-count data, where weak signals are obscured by noise.

### Solution
- **Deep Convolutional Neural Networks**: The authors propose a method based on deep convolutional neural networks to remove noise from low-count data through supervised training. The training data includes pairs of low-count and high-count data obtained from actual experiments.
- **Training Method**: The neural network is trained through supervised learning, using paired low-count and high-count data as input and output. This method effectively extracts weak signals rather than merely reducing noise.
- **Performance Evaluation**: The authors evaluate the denoising effect by comparing the signal-to-background ratio (SRBR) of the denoised low-count data with the high-count data. The results show that the neural network trained with experimental data significantly outperforms the network trained with artificial noise data in terms of denoising effect.

### Main Findings
- **Importance of Experimental Data**: Neural networks trained with experimental data perform better in removing experimental noise, especially when dealing with weak signals. This indicates that while artificial noise data can be used for preliminary testing, the final denoising effect still relies on actual experimental data.
- **Generalization Ability**: The trained neural network not only performs well on XRD data but can also be applied to other experimental techniques, such as resonant inelastic X-ray scattering (RIXS), despite the different data distributions and noise characteristics of these techniques.

### Application Prospects
- **Improving Experimental Efficiency**: By effectively removing noise, researchers can obtain high-quality data in a shorter exposure time, saving a significant amount of experimental time.
- **Expanding Application Range**: This method can be applied to other scattering, spectroscopy, and microscopy techniques, further enriching the training data and improving the robustness and generality of the denoising algorithm.

In summary, this paper proposes a denoising method based on deep convolutional neural networks that can effectively extract weak signals in X-ray diffraction experiments, providing new tools and methods for scientific research.