Minimal droplet shape representation in experimental microfluidics using Fourier series and autoencoders.

This article focuses on a novel two-step, fully reversible method to represent the outer shape of microfluidic droplets in a low-dimensional space. The approach first expresses droplet contours as Fourier series and then compresses the Fourier coefficients using autoencoder neural networks, reducing the dimensionality from 42 real numbers to just two while maintaining high reconstruction accuracy. This minimal representation enables efficient real-time analysis at speeds exceeding 30 frames per second, facilitating applications such as automated droplet generation via reinforcement learning and the study of droplet shape dynamics. The study highlights the advantage of incorporating domain knowledge into machine learning models to achieve significant dimensionality reduction, though it notes that linking the minimal representation to physical droplet properties remains a subject for future research.