Overfitting One-Dimensional convolutional neural networks for Raman spectra identification

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Handheld Raman spectrometers are crucial for in situ substance identification by first responders.
• Current devices rely on extensive reference data, limiting miniaturization and affordability due to memory constraints.

Purpose of the Study:

• To address the limitations of memory and computational power in handheld Raman spectrometers.
• To develop a more efficient and accurate method for real-time spectral matching using machine learning.

Main Methods:

• Utilizing one-dimensional Convolutional Neural Networks (1DCNN) trained on augmented Raman spectra.
• Employing an overfitted 1DCNN model as a substitute for large reference databases.
• Testing the 1DCNN model's performance in identifying pure unknown Raman instances.

Main Results:

• The 1DCNN model significantly reduces the reliance on extensive onboard reference data.
• Experimental results demonstrate high accuracy in identifying unknown Raman spectra from thousands of classes.
• The proposed method alleviates memory size limitations and increases identification speed.

Conclusions:

• 1DCNN offers a viable solution to enhance the capabilities of handheld Raman spectrometers.
• This machine learning approach improves the speed, accuracy, and affordability of in situ substance identification.
• The study paves the way for more advanced and portable spectroscopic analysis tools.