Machine Learning-Enabled Image Analysis of Complex Chemical Mixtures: Synthetic Urine Droplets as a Test System

Dried droplet deposits encode compositional information through evaporation-driven self-organization, offering a potential low-cost sensing modality for chemical analyses. Here, we evaluate the feasibility of inferring solute concentrations and pH from photographic images of dried synthetic urine stains by using automated experimentation and machine learning. A robotic droplet imager (RODI) combined with Latin hypercube sampling was employed to efficiently explore compositional space and generate standardized data sets comprising 70,000 images across two experimental systems: a two-component urea-NaCl model and seven-component synthetic urine mixtures spanning physiologically relevant ranges for urea, NaCl, KCl, creatinine, glucose, a combined salt group, and pH. Each image was reduced to 47 descriptors, capturing spatial morphology and texture. Random Forest regression yielded useful predictive performance for NaCl in the two-component system and identified pH as the most robustly inferred parameter in the complex mixture. However, continuous prediction of individual solute concentrations in the seven-component system remained limited, possibly due to the sparsity of training points across the high-dimensional compositional space. These findings delineate both the current capabilities and information-theoretic constraints of image-based compositional inference for complex chemical mixtures.