Infrared spectroscopy-assisted prediction of impurities in chemicals using machine learning: towards smart self-driving laboratories

Abstract

The changeable nature of chemicals requires constant monitoring. In a self-driving laboratory (SDL), manual revision of chemicals should be replaced by automated compliance checks. Infrared spectroscopy is an SDL compatible tool that can be integrated into a robotic arm to record a spectrum for each compound scheduled to be used in a synthetic transformation. In the case of any changes happening with a chemical, the IR-spectrum would include both signals of the main compound and an impurity, resulting in failure in decoding and further synthesis. Here, a series of complicated IR-spectra of mixtures of compounds was generated to train a neural network. The mixtures were identified based on chemical intuition. An F₁-score of 0.94 for purity determination and 0.97 for functional group determination was achieved for the generated spectra. The trained model was tested using IR-spectra of real mixtures of compounds. An F₁-score of 0.99 for functional group determination was achieved for the real spectra and 0.95 for the purity determination task was observed. In fact, the model was able to determine whether a compound was pure enough for further use or contaminated with the identified substance and should be purified.