Machine Learning-Based Analysis of Molar and Enantiomeric Ratios and Reaction Yields Using Images of Solid Mixtures

Abstract

Visual observations are frequently used for the preliminary evaluation of the chemical contents of mixtures, but their accuracy largely depends on the observer’s experience and intuition, which are difficult to share. Here, we report component ratio prediction using image-based machine learning (ML), which is applicable to the analysis of various solid mixtures, such as mixtures of organics and inorganics, polymorphous crystals, and enantiomers. The trained model with 300 images could predict the sugar/dietary salt weight ratio from an image with 4% error. The ML prediction pipeline was shown to be broadly applicable to polymorphic glycine, d/l-tartaric acid, and four-component systems. As an application demonstration, we also used our ML system to analyze the yield of a solid-state decarboxylation reaction. These results demonstrated that accumulation of researchers’ experience derived from visual information can be shared as trained ML models and used as a quantitative analysis method.