Deep Learning for Reconstructing Low-Quality FTIR and Raman Spectra─A Case Study in Microplastic Analyses

Area of Science:

• Spectroscopy
• Machine Learning
• Environmental Science

Background:

• Automated spectroscopic analysis, particularly for microplastic (MP) particle identification, often yields heterogeneous datasets due to time constraints.
• Low signal-to-noise ratios and baseline artifacts in Fourier transform infrared (FTIR) and Raman spectra hinder accurate identification in automated workflows.
• Manual optimization of spectral post-processing parameters is not feasible in high-throughput automated measurements.

Purpose of the Study:

• To develop and demonstrate a deep-learning method for automated removal of instrumental noise and spectral artifacts.
• To improve the quality of FTIR and Raman spectra acquired in automated, time-limited applications.
• To enhance the reliability of spectral identification in environmental monitoring, specifically for microplastic analysis.

Main Methods:

• Implementation of a simple autoencoding neural network for spectral reconstruction.
• Training the neural network on a dataset with complex spectral distortions, including noise and baseline artifacts.
• Applying the trained network for single-pass artifact removal without spectra-specific parameter tuning.

Main Results:

• The autoencoding neural network successfully reconstructed complex spectral distortions, effectively removing noise and baseline artifacts.
• The method demonstrated high computational efficiency, enabling artifact removal in a single pass.
• The deep-learning approach significantly improved spectral quality for automated analysis.

Conclusions:

• Deep learning offers a powerful and efficient solution for cleaning noisy and artifact-laden FTIR and Raman spectra in automated applications.
• This method holds significant potential for environmental monitoring and other applications requiring rapid, high-volume spectral analysis.
• The developed technique facilitates accurate spectral identification even with challenging, heterogeneous datasets common in automated workflows.