Fig. 2. Overview of ML approaches in microfluidic performance prediction. (A) The performance of a microfluidic device is affected by the flow rates, fluid properties, device geometry, and material. (B) With informed sampling methods, this large design space can be explored experimentally or numerically to create a sufficiently sized dataset for training ML models. (C) Several classes of ML models, such as fully-connected neural networks (fcNN), convolutional neural networks (CNN), and autoencoders can be trained to gain generalizable predictive understanding of microfluidics. (D) Examples of ML-assisted performance prediction in several types of microfluidic devices. (D.i) An autoencoder in conjunction with a fcNN was used to predict droplet stability in tightly packed emulsions passing through a constriction. Figure reproduced from Khor et al., 2019 (ref. 77) with permission from the Royal Society of Chemistry. (D.ii) ML models were used to predict the distribution of lift in broad range of operating conditions, which was then used to estimate particle inertial focusing in multiple cross-sections. Figure reproduced from Su et al., 2021 (ref. 15) with permission from the Royal Society of Chemistry. (D.iii) CNNs were used to predict the solute concentrations and flow rates at the three outlets of a 2-inlet randomly designed grid micromixer. Figure reproduced from Wang et al., 2021 (ref. 78) with permission from the Royal Society of Chemistry.