Infrared spectroscopy-based zero-shot learning for identifying reaction intermediates in unseen systems

Identifying reaction intermediates is a critical component of elucidating the mechanisms of chemical reactions. Spectroscopic techniques are instrumental in this identification process. The capacity of AI to establish the correlations between spectra and chemical substances renders it a suitable tool for spectra analysis. However, the presence of limited data, or even the absence of data, is a common occurrence for intermediates with transient characteristics. This poses a significant challenge in meeting the data requirements of AI models. Herein, we propose a generalizable machine learning model that utilizes zero-shot learning to identify chemical reaction intermediates by their spectra in unseen catalytic systems. Using SHapley Additive exPlanations (SHAP) analysis and visual dimensionality reduction, it was determined that the model's superior generalizability is attributable to its capacity of learning common patterns of spectra-intermediates, effectively mapping disparate catalytic systems to analogous digital spaces. Therefore, the model can be directly used without fine-tuning parameters for unseen systems even in the presence of noise or solvents. This work demonstrates the application of zero-shot learning in machine learned spectroscopy and illustrates the prediction mechanism, providing a perspective for interpretable and robust cross-system prediction. It lays a solid foundation for the application of spectral descriptors in real reaction systems and the comprehension of chemical reaction mechanisms.