Fig. 4: Models’ performance in predicting stable single-core and unstable DE droplet generation regimes.

a DE generation was modeled as two events of droplet generation at FF1 (aqueous-in-oil) and FF2 (oil-in-aqueous); the threshold for unstable DE generation (multiple core or missing core) was set to a generation rate difference (GRD) of 15%. b Comparison of predicted and calculated GRD at FF1 and FF2 for (i.) the neural network, (ii.) boosted decision trees, and (iii.) consensus model over the 197 stable datapoints. Green boxes indicate regions with predicted GRD <15% and experimentally stable DEs; predictions are shown for a single representative model. (iv.) The effect of GRD threshold on the true positive rate of stable DE generation predictions (i.e., correctly predicted to be stable) for different models; values represent average accuracies and error bars indicate two standard deviations over 10 randomized training sessions. c Comparisons between errors in model-predicted generation rates at FF1 and FF2. Markers show comparisons for a single representative model and dashed line indicates 1:1 line. d Comparisons between observed mode of instability vs. predicted GRD for (i.) the neural network, (ii.) boosted decision trees, and (iii.) the consensus model. Correct predictions appear in green shaded areas, incorrect predictions appear in red shaded areas, and GRDs predicted to lead to stable droplets are indicated by lighter shading. (iv.) Plots quantifying accuracy in predicting the mode of instability and (v.) the true negative rate (i.e., percentage of unstable DE generation data that were correctly predicted to be unstable) for different GRD thresholds. Plot values indicate average accuracy of predictions and error bar’s total length indicates two standard deviations over 10 randomized training sessions. Source Data are provided as a Source Data file.