High uniformity hohlraum designs are essential for achieving high yield in indirect-drive inertial confinement fusion experiments. However, only recently has there been motivation toward spherical-based hohlraums.
The PEPR hohlraum is a novel geometry for the OMEGA 60-beam laser. This hohlraum would provide excellent implosion uniformity and promise insight towards spherical hohlraum performance.
View Paper (Physics of Plasmas)
Fiber networks are prevalent throughout nature, from cartilage to zebrafish embryos. These biological tissues demonstrate unique mechanical properties, such as its reponses to stress. For example, top layer of cartilage is up to 10-100 times more compliant than the deeper layers.
I wrote simulations to model the mechanical responses of these fiber networks. After applying a strain onto the network, highly-optimized (GPU-accelerated, sparsified, vectorized, pre-conditioned) numerical methods find minimal energy state of the networks. The code is written in a modular-fashion and generalized enough in that it can be applied to both 2-D and 3-D lattices.
As part of my Functional Analysis final project, I extended previous works looking at gradient flow under the Wasserstein-2 metric. I consider a new gradient flow update which imitates a descent method typically used for structural relaxation.
I define new modifications to Hamiltonian dynamics to incorporate both damping and "steering" for the momentum updates. The following GIFs show the accelerated method with KL divergence as the energy functional.
View Repo or View Writeup
An ignition-scale laser system capable of comparing both direct-drive and indirect-drive with high uniformity (on the same target chamber) does not yet exist. We propose a port configuration that supports symmetric direct-drive implosions with minimal adjustments of the beam pointings
I developed a 3D viewfactor code to simulate capsule uniformity during indirect-drive implosions. The program was written in FORTRAN (and using OpenMP for multiprocessing). I used the code to compare hohlraum geometries as well as test the possibilities of new target chamber designs.
View Presentation Abstract
Quantifying and drawing outlines of cells in fluorescence microscopy images is especially important to studying cellular responses and to biological sciences as a collective. However, the process of manually segmenting cells is tedious and challenging, especially when tracking videos of cells. A video of small sample of impacted articular cartilage can have as many as 5000 cells across 1000+ time points.
An implementation of Mask R-CNN was trained on images/frames of impacted cartilage. The initial weights were pre-trained on the COCO dataset to support transfer learning, so only the head layers were trained. Due to the noisiness of the dataset and limited training time, the model was not robust enough to perform segmentation to a high certainty. However, the results suggest that much of noise and inaccuracies within the pre-labeled dataset were not as present within the model.
During high school, I developed several softwares at CCRC of University of Rochester. The first being QTClock (shown on the right).
QTClock obtained data from ECG recordings and plotted both the QT interval along with a color gradient corresponding to the patient’s intake drug concentration in a clock-like graph. I also developed an online calculator that assesses the absolute risk of life-threatening cardiac events in patients with long QT syndrome.
View Paper