Machine Learning-Assisted Phase Diagram Determination in Aqueous Two-Phase Systems

Accurate construction of phase diagrams for aqueous two-phase systems (ATPSs) is essential for identifying two-phase compositions for downstream applications. However, current experimental strategies based on turbidimetric titration are labor-intensive and hindered by inherent inaccuracies. In this study, we developed an automated methodology integrating optical microscopy and machine learning to detect microscopic phase separation and define phase boundaries. Our pipeline automatically analyzes thousands of images using a state-of-the-art segmentation model (SAM 2) to identify and quantify emulsion droplets even when droplet edges are dim or occluded. Once detected, supervised machine learning is used on the droplet features to guide the classification into one- or two-phase regions. Using this approach, we identified two-phase regions for an ATPS consisting of 10 kg/mol poly(ethylene glycol) and 10 kg/mol dextran, a phase diagram of which is hard to construct accurately using the conventional titration method due to slow phase separation kinetics. Overall, this strategy provides a robust and efficient platform for automated phase diagram construction in ATPSs, with potential for broad applications in phase-separating systems.