Source author record

Vidya Ganapati

Vidya Ganapati appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

physics.optics Distributed, Parallel, and Cluster Computing Machine Learning

Catalog footprint

What is connected

3works

3topics

4close collaborators

Actions

Connect this record

Open graph Browse works

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2021arXiv

HYPPO: A Surrogate-Based Multi-Level Parallelism Tool for Hyperparameter Optimization

We present a new software, HYPPO, that enables the automatic tuning of hyperparameters of various deep learning (DL) models. Unlike other hyperparameter optimization (HPO) methods, HYPPO uses adaptive surrogate models and directly accounts for uncertainty in model predictions to find accurate and reliable models that make robust predictions. Using asynchronous nested parallelism, we are able to significantly alleviate the computational burden of training complex architectures and quantifying the uncertainty. HYPPO is implemented in Python and can be used with both TensorFlow and PyTorch libraries. We demonstrate various software features on time-series prediction and image classification problems as well as a scientific application in computed tomography image reconstruction. Finally, we show that (1) we can reduce by an order of magnitude the number of evaluations necessary to find the most optimal region in the hyperparameter space and (2) we can reduce by two orders of magnitude the throughput for such HPO process to complete.

preprint2014arXiv

Intermediate Mirrors to Reach Theoretical Efficiency Limits of Multi-Bandgap Solar Cells

Creating a single bandgap solar cell that approaches the Shockley-Queisser limit requires a highly reflective rear mirror. This mirror enhances the voltage of the solar cell by providing photons with multiple opportunities for escaping out the front surface. Efficient external luminescence is a pre-requisite for high voltage. Intermediate mirrors in a multijunction solar cell can enhance the voltage for each cell in the stack. These intermediate mirrors need to have the added function of transmitting the below bandgap photons to the next cell in the stack. In this work, we quantitatively establish the efficiency increase possible with the use of intermediate selective reflectors between cells in a tandem stack. The absolute efficiency increase can be up to ~6% in dual bandgap cells with optimal intermediate and rear mirrors. A practical implementation of an intermediate selective mirror is an air gap sandwiched by antireflection coatings. The air gap provides perfect reflection for angles outside the escape cone, and the antireflection coating transmits angles inside the escape cone. As the incoming sunlight is within the escape cone, it is transmitted on to the next cell, while most of the internally trapped luminescence is reflected.

preprint2013arXiv

Light Trapping Textures Designed by Electromagnetic Optimization for Sub-Wavelength Thick Solar Cells

Light trapping in solar cells allows for increased current and voltage, as well as reduced materials cost. It is known that in geometrical optics, a maximum 4n^2 absorption enhancement factor can be achieved by randomly texturing the surface of the solar cell, where n is the material refractive index. This ray-optics absorption enhancement limit only holds when the thickness of the solar cell is much greater than the optical wavelength. In sub-wavelength thin films, the fundamental questions remain unanswered: (1) what is the sub-wavelength absorption enhancement limit and (2) what surface texture realizes this optimal absorption enhancement? We turn to computational electromagnetic optimization in order to design nanoscale textures for light trapping in sub-wavelength thin films. For high-index thin films, in the weakly absorbing limit, our optimized surface textures yield an angle- and frequency-averaged enhancement factor ~39. They perform roughly 30% better than randomly textured structures, but they fall short of the ray optics enhancement limit of 4n^2 ~ 50.