Deep learning-based statistical noise reduction for multidimensional spectral data

Abstract

In spectroscopic experiments, data acquisition in multi-dimensional phase space may require long acquisition time, owing to the large phase space volume to be covered. In such a case, the limited time available for data acquisition can be a serious constraint for experiments in which multidimensional spectral data are acquired. Here, taking angle-resolved photoemission spectroscopy (ARPES) as an example, we demonstrate a denoising method that utilizes deep learning as an intelligent way to overcome the constraint. With readily available ARPES data and random generation of training datasets, we successfully trained the denoising neural network without overfitting. The denoising neural network can remove the noise in the data while preserving its intrinsic information. We show that the denoising neural network allows us to perform a similar level of second-derivative and line shape analysis on data taken with two orders of magnitude less acquisition time. The importance of our method lies in its applicability to any multidimensional spectral data that are susceptible to statistical noise.

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

“…Small-angle neutron and X-ray scattering data [26] would be an obvious path to extend the training data. Spectroscopies [27,28] and microscopies [29,30] such as angle-resolved photoemission electron spectroscopy [31] and transmission electron microscopy [32,33] data could also help expanding the amount and variety of training data. Data from astronomy and medical scanners would be another way to train the networks on a wider spectrum of noise sources.…”

“…Loss function: During each training epoch, the performance of the neural networks is determined by comparing the denoised output with the high-count frame. The used loss function L is given by a combination of mean absolute error (MAE) and multiscale structural similar-ity (MSSIM) [27,37] L = (1 − α)L MAE + αL MSSIM with α = 0.7 where the MAE between two images x and y is given by the sum over all pixels i MAE(x, y) =…”

“…Loss function: During each training epoch, the performance of the neural networks is determined by comparing the denoised output with the high-count frame. The used loss function L is given by a combination of mean absolute error (MAE) and multiscale structural similarity (MSSIM) [27,37] L = (1 − α)L MAE + αL MSSIM with α = 0.7 where the MAE between two images x and y is given by the sum over all pixels i MAE(x, y) = where µ x,y , σ x,y and σ xy are the mean, standard deviation and covariance of x and y. K 1 and K 2 are small scalar constants ≪ 1 and L defines the dynamic range of pixel values. This measure iteratively downsamples the input image by a factor of two for every scale.…”

A machine learning route between band mapping and band structure

“…Therefore, their use is limited to selected momentum locations determined heuristically from physical knowledge of the materials and the experimental settings. Image processing-based methods apply data transformations to improve the visibility of dispersive features [15][16][17][18]. They are more computationally efficient and can operate on entire datasets, yet offer only visual enhancement of the underlying band dispersion.…”

Transfer Learning Application of Self-supervised Learning in ARPES

“…For example, a deep layer of convolutional neural network (ConvNet) is trained to denoise ARPES data [28]. Afterwards, there are efforts to obtain how the bandstructure calculation based on the ARPES data [29,30], where reference [30] provides additional feature of obtaining the result even through a noisy data (simulated noise).…”