Diffusion plays an important role in many physical and biological processes. Understanding the physical mechanism underlying diffusion is crucial to study transport in complex systems. The development of single particle tracking (SPT) has revealed important information about particle dynamics that cannot be obtained by observing ensemble averages. The study of single diffusion trajectories is a challenging task due to the limited information contained in them and their stochastic nature. Recently, machine learning methods have been successful in characterizing diffusion. In this work we propose unsupervised learning as means to study anomalous diffusion at a theoretical level and as a tool to characterize individual trajectories. We developed an autoencoder, a neural network which is able to learn a representation of input trajectories in a lower dimensional space which we refer to as latent space. We trained our neural network with simulated diffusion trajectories and analyzed the latent space of the autoencoder to obtain information about their nature. Additionally, we used the autoencoder to compare theoretical diffusion models with anomaly detection. With the help of supervised learning, we studied the transformations that can be applied by an autoencoder to modify the original data and obtain new information.
Major project supervisor
Minor project supervisor