Unsupervised machine learning-based image recognition of raw IR spectra: Toward chemist-like chemical structural classification and beyond numerical data

Recent advances in artificial intelligence have significantly improved spectral data analysis. In this study, we used unsupervised machine learning to classify chemical compounds based on infrared (IR) spectral images, without relying on prior chemical knowledge. The potential of machine learning for chemical classification was demonstrated by extracting IR spectral images from the Spectral Database for Organic Compounds and converting them into 208620-dimensional vector data. Hierarchical clustering of 230 compounds revealed distinct main clusters (A–G), each with specific subclusters exhibiting higher intracluster similarities. Despite the challenges, including sensitivity to spectral deviations and difficulty of distinguishing delicate chemical structures in spectra with low transparency in the fingerprint area, the proposed image recognition approach exhibits good potential. Both principal component analysis (PCA) and k-means clustering produced similar results. Furthermore, the method demonstrated high robustness to noise. The Tanimoto coefficient was used to evaluate the molecular similarity, providing valuable insights. However, some results deviated from chemists’ intuitions. The study also highlighted that the scaling composition formulas and molecular weights did not affect the classification results because high-dimensional features dominated the process. A comparison of the clustering results obtained from molecular fingerprints, using the adjusted Rand index as a metric, indicated that the image data provided better classification performance than numerical data of the same resolution. Overall, this study demonstrates the feasibility of using machine learning with IR spectral image data for chemical classification and offers a novel perspective that complements traditional methods, although the classifications may not always align with chemists’ intuitions. This approach has broader implications for fields such as drug discovery, materials science, and automated spectral analysis, where handling large, raw spectral datasets is essential.