Unsupervised Raw IR Spectra Images Recognition: Toward Chemist-Like Chemical Structure Classification

The recent advancements in artificial intelligence have greatly improved spectral data analysis. Here, we explored using unsupervised machine learning to classify chemical compounds based on IR spectrum images without relying on prior chemical knowledge. The research demonstrated the potential of machine learning in chemical classification by extracting IR spectral images from the SDBS database and converting them into 218196-dimensional vector data. The hierarchical clustering of 227 compounds revealed distinct main clusters (A-G), each with specific subclusters showing higher intra-cluster similarity. Despite some challenges, such as sensitivity to spectral deviations and difficulty distinguishing delicate chemical structures in spectra with low transparency in the fingerprint area, the approach showed promise. The Tanimoto coefficient was used as a metric for molecular similarity, providing valuable insights, though sometimes diverging from chemist intuition. The study also highlighted that scaling composition formulas and molecular weights did not affect the classification results, as the high-dimensional features dominated the process. Overall, the research demonstrated the feasibility of using IR spectral image data in machine learning for chemical classification, offering a novel perspective that complements traditional methods, even though some classifications may not always align with chemist intuition.