Fig. 1: Closed loop navigation of condensate phase diagrams.

The workflow is constituted by three parts: (I) condensate formulation, where samples are automatically prepared, (II) confocal microscopy and sample classification, for characterization, and (III) active machine learning, that learns from the collected data and suggests the next experiments. A Condensate microparticles are formed by mixing cationic and anionic polypeptides, resulting in phase-separated micron-sized droplets. B Schematic representation of the robotic pipetting platform with 16 flexible deck slots. C Formulations are prepared in a conical PCR plate, using contactless dispensing with volume tracking. A custom touch-tip functionality follows a touch-point trajectory to ensure accurate dispensing. D Confocal imaging is performed using dynamic Z-stack acquisition. E Example segmentation of representative confocal microscopy data using automated binary Yen-thresholding for particle detection in each Z-plane. F The optimal Z-plane is selected based on the largest detected area, corresponding to the slice that is best in focus. G Samples with 12 or more particles are labeled as phase-separated (condensates), while those below the threshold are labeled as non-condensates. H Experimentally validated data points are incorporated into the machine learning algorithm for training. I The model predicts a phase diagram based on the acquired experimental data. J The model then guides the selection of new formulations, restarting the automation cycle at (A).