Differential Dynamic Microscopy Enhanced with Convolutional Neural Networks

Differential dynamic microscopy (DDM) has been widely used to analyze the dynamics of bacteria, colloidal particles, gels, and other soft matter systems. With this Fourier-space image analysis technique, one typically needs 100s to 1000s of images in a time series to accurately quantify dynamics. When fewer images are used for DDM analysis, the output can be too noisy to reliably quantify the dynamics. Here, we employ a convolutional neural network (CNN) to denoise DDM data. With this machine learning approach, we can obtain accurate values for the characteristic decay times of density fluctuations in colloidal suspensions using only 10s of frames. The ability to accurately perform DDM with short durations of imaging data enables the study of samples in which the dynamics are quickly changing. We demonstrate this approach of using DDM with machine learning to quickly perform microrheology measurements on colloidal particles throughout a sample slide with spatially varying properties, highlighting the high-throughput microrheology enabled here.