Optimization of thin film photovoltaic structures using non-parametric regression

Shima Hajimirza

Optimization of thin film photovoltaic structures using non-parametric regression

Shima Hajimirza

Texas A&M University, USA

: J Nucl Ene Sci Power Generat Technol

Abstract

Designing nano-scale materials with high external quantum efficiency is a complicated inverse design problem requiring the optimization of a utility function derived from the solution of an underlying dynamic system governed by electromagnetic Maxwell and electronic carrier mobility equations. Such design usually consists of layering or texturing the semiconductor material with metallic reflectors, nano-particles or wave-guides, transparent conductive layers, surface roughness and other simultaneous light trapping/career mobility improvement mechanisms and therefore lies within an extremely vast search space. The design problem is consequently very time consuming and challenging, and offers little intuitive understanding of the intertwined effects of joint variable deltas. These types of design problems can be handled more practically via the use "meta-model" optimizations. The idea of such approaches is that the utility function is locally or globally fit with an approximate meta-model based on past search points or carefully designed experiment points. The meta-model is then used as a guide to carry out a simpler optimization closely resembling that of the original design, or to recommend valuable candidate points. In the current work, we report on using a variety of such meta-model approaches to design highly efficient thin film silicon solar cells with multi-layered front and back coatings. We apply different meta-model optimization techniques to the proposed problem and compare the results to standard numerical optimization algorithms. In particular, we propose two novel techniques that have their roots in machine-learning applications; one based on a regularized linear local fit and the other one based on a global regression-tree fit, and we demonstrate that the with the help of these smart techniques we can indeed expedite nano-material search problems to a great extent.