Fig. 5: Generalization of machine learning models to fluids, geometries, and device materials not included within the comprehensive dataset.

a References for previously published and additionally generated experimental data and schematics showing device geometries, surface properties, and fluid compositions used to generate SE (two left devices) and DE droplets (two right devices). b Comparison between observed diameters and those predicted by a previously proposed scaling law and 3 different trained models for a single representative training session. Dashed lines indicate 1:1 identity line, and annotation denotes the average mean absolute percentage error over 15 randomized training sessions. c Quality metrics model performance in predicting droplet diameter. Bars indicate average performance across 15 randomized training sessions; error bars' total length denotes two standard deviations. Magenta color highlights the top-performing model. Source Data are provided as a Source Data file.