Unveiling Spectrum–Structure Correlation in Vibrational Spectroscopy: Task-Driven Deep Learning Classification Balancing Global Fusion and Local Extraction

Spectrum-structure correlation is crucial to identify and quantify chemicals, in which classification of mixtures and identification of functional groups are two central tasks. Deep learning-driven algorithms have made significant strides to these two tasks. However, many of these algorithms are merely adaptations of models originally designed for computer vision applications. As a result, the models often suffer from either low accuracy or limited generality when applied to spectral data due to the overlooked inherent limitations in feature richness and volume of spectral data. Here, in light of the distinctive difference in the attention of global and local information in spectral data between these two tasks, we developed contrapuntally two CNN-based algorithms, incorporating multiscale convolution and attention mechanism, to address the unique requirements of each task. It was found that the lightweight CNN-Peak algorithm is favored for the classification of a mixture, a type of single-label task, in which the feature fusion of global information is more important. Meanwhile, the more complex ResNet-ResPeak algorithm is ideally suited for the identification of functional groups, a type of multilabel task, in which the feature extraction of local information takes precedence. The task-oriented, conceptual design of deep learning algorithms not only enhances the efficacy and accuracy of spectrum-structure correlation analysis but also feeds back to achieve a more rigorous experimental design and implementation, forming a closed loop of AI for Science.