The instrument shines an infrared beam at a sample, and molecules respond with their own absorption pattern, like bells of different tones. This pattern, the spectrum, holds information about the chemical composition. But to read it, a specialist usually consults a library of reference records, which is slow and not always possible.
Infrared spectroscopy is a workhorse tool for chemists, but interpreting spectra remains a bottleneck. The new model UltraIR learns differently: first, it goes through 60 million synthetic spectra generated according to the laws of physics, learning to reconstruct them, compare molecular fingerprints, and predict functional groups. Then it is fine-tuned on a small set of real data, and it already handles real-world tasks.
The model was tested on various tasks: identifying functional groups, determining molecular structures, predicting properties, quantitative analysis of mixtures, classifying bacteria, determining the geographic origin of medicinal herbs, recognizing microplastics, and soil properties. In all cases, UltraIR outperformed conventional machine learning methods and specialized neural networks.
The key advantage is knowledge transfer. A model trained on one infrared spectrometer can be applied to data from another instrument or laboratory without additional training. This paves the way for fast and inexpensive analyses of complex samples in real-world settings.
The infrared spectrum of a substance is like a fingerprint: two different compounds with the same pattern are practically nonexistent.