Development of an automated high-throughput screening platform for the dynamic phase separation analysis of microemulsion systems with AI image recognition

Increasing efforts are undertaken to develop new sustainable production processes, and homogeneous catalysis offers many advantages regarding selectivity and energy efficiency in new chemical production routes. A major factor often limiting the application of homogeneous catalysis is the retention of valuable catalysts. One promising option to introduce superior reaction performance and catalyst recovery in organic reactions is the use of water-soluble catalysts in aqueous reaction media with surfactants. However, these surfactant-based microemulsion systems (MES) exhibit a complex phase separation behavior that is dependent on various parameters such as temperature and component concentrations, while its prediction is currently not possible due to the complex thermodynamics. The characterization of the phase behavior hence requires extensive and time-consuming experimental investigation due to a lack of fundamental modeling approaches. To facilitate the acquisition of experimental data, this contribution presents the development of a high-throughput screening platform for dynamic phase separation analysis with an automated experimental procedure, AI analysis of separation images, and automated result data handling. The platform enables a fast characterization of MES separation behavior, which is required for process development and operation. The functionalities of the screening platform are demonstrated in a case study for the hydroformylation of decene. The image detection is performed with a Mask R-CNN model achieving a ±1.5% accuracy in phase height detection with a classification confidence threshold of 96%. The new setup enables a fast evaluation of over 722 measurement runs each with a different combination of separation temperature and mixture composition that only required at total of 38 manual dosing steps. The gathered data is also used to derive a correlation for a soft-sensor with interpretable machine learning, enabling online insights into otherwise inaccessible process variables in an MES plant and enabling its operability.