Fig. 5. Outlook for machine learning in microfluidics. (A) To effectively learn microfluidic performance across the field, community repositories are needed consisting of standardized designs, fabrication protocols, and performance metrics together with predictive models previously trained on standardized datasets. These datasets and models can be retrieved by researchers, and adapted to new applications through transfer learning with additional smaller datasets. Sharing the updated models on community repositories completes a positive feedback cycle to continuously extend the predictive understanding of microfluidics to new components and applications. (B) Once the performance of enough components are accurately predicted, they can be integrated into an ML-guided “robot scientist” that can route together different components for fully-automated hypothesis testing and optimization. (C) Such a system has been successfully implemented in automated synthesis of organic compounds using millifluidic flow modules. From Coley et al., 2019.¹¹³ Reprinted with permission from AAAS.