Deep learning-based statistical noise reduction for multidimensional spectral data.

Area of Science:

• Spectroscopy
• Data Science
• Materials Science

Background:

• Multidimensional spectroscopic experiments often require extensive data acquisition times due to large phase space volumes.
• Limited acquisition time poses a significant constraint for obtaining high-quality multidimensional spectral data.

Purpose of the Study:

• To demonstrate a deep learning-based denoising method to overcome acquisition time limitations in spectroscopic experiments.
• To preserve intrinsic data information while effectively removing noise.

Main Methods:

• Utilizing a denoising neural network trained on readily available angle-resolved photoemission spectroscopy (ARPES) data and randomly generated training datasets.
• Ensuring the neural network was trained without overfitting.

Main Results:

• The denoising neural network successfully removed noise from ARPES data while preserving essential information.
• Analysis of denoised data showed comparable results to undenoised data acquired with two orders of magnitude longer acquisition time.
• Demonstrated the ability to perform second-derivative and line shape analysis on significantly reduced datasets.

Conclusions:

• Deep learning offers an intelligent solution to the challenge of long acquisition times in multidimensional spectroscopy.
• The developed denoising method is broadly applicable to various multidimensional spectral data susceptible to statistical noise.
• This approach significantly accelerates experimental data acquisition without compromising analytical quality.