Improved quantification of colloidal dynamics with machine learning and simulations

Techniques like dynamic light scattering and particle tracking are often used to quantify the dynamics or rheological properties of soft matter systems or complex fluids. A more recent technique that combines features of optical microscopy and light scattering is differential dynamic microscopy (DDM). This technique utilizes microscopy-acquired real-space videos to calculate correlation functions using a framework similar to light scattering methods. DDM has been used to measure the diffusion of colloids or nanoparticles, the dynamics of gels, and the fluctuations of cytoskeleton networks driven by molecular motors. In most of these cases, 1000s of image frames are necessary to analyze the dynamics accurately with DDM. To overcome the need to acquire large amounts of imaging data, we use a convolutional neural network (CNN) to denoise DDM data. With this machine learning approach to DDM, we can accurately quantify motion using imaging data acquired in significantly less time. Training of the convolutional neural network can involve using vast amounts of experimentally acquired data across a range of conditions. More efficiently, we can train the CNN using simulated video microscopy data. We show that models trained on simulations work well on experimental data. With this approach, we quantify dynamics that are quickly evolving in time and demonstrate how high-throughput measurements of rheological properties can be performed.