Deep learning for 'artefact' removal in infrared spectroscopy.

Area of Science:

• Spectroscopy
• Data Science
• Materials Science

Background:

• Infrared (IR) spectra of heterogeneous media are distorted by optical phenomena like scattering and reflection.
• These distortions, including baseline shifts and band changes, obscure molecular information and invalidate the Beer-Lambert law.
• Accurate interpretation of IR spectra requires removing these optical artefacts to recover pure sample absorbance.

Purpose of the Study:

• To develop and validate a deep learning approach for removing optical artefacts from IR spectra.
• To recover the true absorbance of a sample from its measured apparent absorbance.
• To assess the generalization performance of the proposed method on diverse datasets.

Main Methods:

• A 1-dimensional U-Net deep convolutional neural network (1D U-Net) was employed for artefact removal.
• A simulated dataset of apparent absorbance and true absorbance pairs for poly(methyl methacrylate) (PMMA) spheres was generated using Mie theory.
• The 1D U-Net was trained on augmented simulated data and validated using independent simulated and experimental datasets.

Main Results:

• The 1D U-Net successfully retrieved pure absorbance, even when optical artefacts dominated the measured spectra.
• High accuracy in artefact removal was confirmed by the hit-quality-index (HQI) comparing corrected and true absorbance.
• The network demonstrated good generalization capabilities, performing well on data with different artefact patterns than the training set.

Conclusions:

• Deep learning, specifically 1D U-Net, offers a powerful solution for correcting optical artefacts in IR spectroscopy.
• The method significantly improves the interpretability of IR spectra from complex heterogeneous samples.
• The validated approach provides a reliable tool for accurate molecular analysis in materials science and beyond.