Ruoshi Liu
“Detecting Topological Defects in Confined 2D Active Nematics with Deep Learning”
Ruoshi Liu, Mike Norton, Pengyu Hong, Seth Fraden
Abstract
An active nematic suspension composed of microtubule bundles driven by ATP-powered kinesin clusters continuously flows in a turbulent-like manner and produces motile topological defects. By confining the suspension within a circular disk, a total topological charge on the system is enforced requiring, at a minimum, two +1/2 defects. Under certain confinement conditions, these two defects co-rotate around each other in the well. This circulating state is constantly disturbed by the nucleation of new defects near the boundary followed by annihilations of old defects. In order to understand these dynamics more fully, defects must be identified reliably. The previous tracking method required identification of the nematic director field before disclinations could be found. This process worked well only under ideal experimental and imaging conditions but failed in many cases where human observers could still identify defects.
To increase the reliability of defect tracking, we implemented the current state-of-the-art object detection algorithm YOLO. We labeled 9000 frames of experimental videos and used this library of defects to train a 24-layer convolutional neural network with the deep learning library Pytorch. Network training and object identification both utilize GPU computing resources. The detector succeeds in detecting scenarios that appear more often in the experimental videos such as two defects co-rotating around each other but has difficulty with detecting anomalous cases such as nucleation events. This detector is currently being optimized by 1) image augmentation; 2) increasing the number of classes for better specificity; and 3) increasing the depth of the neural network.
In the next phase, a deep Recurrent Neural Network will be constructed to predict the nucleation events, which will shed light on the dynamics of the confined active 2D nematics.
Support: SMURF (Summer MRSEC Undergrad Research Fellowship)