Dense Convolutional Neural Network for Identification of Raman Spectra.

Area of Science:

• Spectroscopy
• Artificial Intelligence
• Cloud Computing

Background:

• Cloud computing and deep learning enable widespread intelligent applications.
• Raman spectra identification can be performed in the cloud, reducing reliance on terminal instruments.
• Environmental interferences can significantly decrease the accuracy of Raman spectra identification algorithms.

Purpose of the Study:

• To propose a deep learning algorithm for accurate cloud-based Raman spectra identification.
• To address the challenge of decreased identification accuracy due to environmental interferences.
• To enhance the robustness and adaptability of Raman spectra identification in diverse conditions.

Main Methods:

• A deep learning algorithm based on a Dense convolutional neural network (CNN) with over 40 layers was developed.
• The Dense network features a feed-forward connection in its Dense blocks, mitigating gradient issues and enhancing feature reuse.
• A database of 1600 Raman spectra from 32 liquid chemicals, including interfered spectra, was created for testing.

Main Results:

• The proposed Dense CNN demonstrated superior accuracy and robustness compared to other CNN-based algorithms.
• The algorithm achieved a weighted accuracy of 99.99% on the custom database across 50 repeated training and testing sets.
• The Dense network also showed good performance when tested on the RRUFF database.

Conclusions:

• The developed Dense network-based approach significantly advances cloud-enabled Raman spectra identification.
• The method effectively mitigates noise interferences, ensuring precise identification even in complex environments.
• This approach offers improved accuracy and adaptability for various Raman spectra identification tasks.