Deep Learning-Assisted Multivariate Analysis for Nanoscale Characterization of Heterogeneous Beam-Sensitive Materials

Area of Science:

• Materials Science
• Data Science
• Nanotechnology

Background:

• Nanoscale characterization employs energetic probes that damage beam-sensitive materials.
• Low-dose probes minimize damage but reduce data signal-to-noise ratio (SNR).
• Existing methods for denoising, like principal component analysis (PCA), face challenges with accurate component separation.

Purpose of the Study:

• Optimize and validate multivariate analysis techniques for low-dose nanoscale characterization data.
• Develop an automated method for accurate separation of signal and noise components.
• Enhance the extraction of meaningful information from beam-sensitive materials.

Main Methods:

• Applied principal component analysis (PCA) and nonnegative matrix factorization for postprocessing.
• Investigated PCA component separation using scree plots and manual selection.
• Developed and validated a deep learning-based neural network for automated PCA component classification.

Main Results:

• PCA demonstrated superior performance for data denoising in low-dose nanoscale characterization.
• Manual separation of PCA components improved SNR by an order of magnitude but was time-consuming.
• The deep learning method achieved >99% accuracy in classifying PCA components at approximately 2 components/s.

Conclusions:

• Multivariate analysis, particularly PCA, is effective for denoising low-dose nanoscale data.
• Automated component classification using deep learning overcomes limitations of manual selection.
• The combined approach enables more robust and efficient characterization of beam-sensitive nanomaterials.