Unveiling Spectrum-Structure Correlation in Vibrational Spectroscopy: Task-Driven Deep Learning Classification...

Area of Science:

• Chemistry
• Artificial Intelligence
• Spectroscopy

Background:

• Spectrum-structure correlation is vital for chemical identification and quantification.
• Current deep learning models, often adapted from computer vision, struggle with spectral data's unique characteristics.
• This leads to suboptimal accuracy and generalizability in chemical analysis tasks.

Purpose of the Study:

• To develop specialized deep learning algorithms for chemical spectrum-structure correlation.
• To address the distinct information needs (global vs. local features) of mixture classification and functional group identification.
• To improve the accuracy and applicability of AI in chemical analysis.

Main Methods:

• Developed two Convolutional Neural Network (CNN)-based algorithms: CNN-Peak and ResNet-ResPeak.
• Incorporated multiscale convolution and attention mechanisms tailored to spectral data.
• Designed algorithms with distinct architectures to leverage global (CNN-Peak) or local (ResNet-ResPeak) feature extraction.

Main Results:

• CNN-Peak, a lightweight model, excels at mixture classification (single-label tasks) by effectively fusing global spectral information.
• ResNet-ResPeak, a more complex model, is superior for functional group identification (multilabel tasks) due to its emphasis on local feature extraction.
• Both algorithms demonstrate enhanced efficacy and accuracy compared to general computer vision adaptations.

Conclusions:

• Task-specific deep learning algorithm design is crucial for optimizing spectrum-structure correlation.
• The developed algorithms, CNN-Peak and ResNet-ResPeak, offer improved performance for specific chemical analysis tasks.
• This work establishes a closed-loop system of AI for Science, enhancing both algorithmic development and experimental design.