Flow-Induced Polarization of Topological Defect Multipoles in Confined Nematic Liquid Crystals

Flow-Induced Polarization of Topological Defect Multipoles in Confined Nematic Liquid Crystals SI Code Code for convolutional neural network-based machine learning To quantitatively classify fluid flow characteristics from polarized LC micrographs, we developed a convolutional neural network (CNN) utilizing the VGG-16 architecture. The CNN was trained to predict fluid flow velocity and azimuthal angle based on input micrographs (Fig. 4A and S3). The training pipeline incorporated efficient data handling via cached TensorFlow datasets and employed mixed-precision optimization to enhance computational efficiency. Training outputs included checkpoints, model weights, accuracy and loss curves, and confusion matrices. Model convergence was achieved within 27 epochs (~27 min per run) on an NVIDIA A100 GPU, demonstrating both the computational efficiency and robust predictive capability of this approach. Code for convolutional neural network-based Grad-CAM visualization To enhance interpretability of the CNN predictions, we implemented Grad-CAM visualization techniques. Given a trained VGG-16 CNN model, this script generates Grad-CAM visualizations for specified input images. The visualizations highlight regions within the LC micrographs contributing significantly to the prediction outcomes, computed specifically for the five convolutional block layers of the network. Each visualization undergoes a 99.5-percentile contrast stretch for clarity and generates a heatmap representation, facilitating detailed analysis (Fig. 4C).