Title

Parallelization Techniques of a Multi-Objective Evolutionary Algorithm for Quantum Cascade Laser Design Optimization

Poster Number

12

Lead Author Affiliation

Engineering Science

Lead Author Status

Masters Student

Second Author Affiliation

Electrical and Computer Engineering

Second Author Status

Faculty

Introduction

Evolutionary computation is a set of optimization algorithms that use the basic processes of Darwinian nature: selection, breeding, mutation, inheritance, etc. to make decisions about a population based on objectives. This process is iterative and the intended outcome is progress, based on the objectives, in the population. In this work, the techniques of evolutionary computation, specifically multi-objective evolutionary algorithms (MOEAs), are used to search the massive design space of the diagonal transition quantum cascade laser (QCL) Quantum cascade lasers have become the standard for high-efficiency emission sources in the mid-infrared (mid-IR) and THz spectrums. Applications such as infrared counter measures, trace gas detection, and free-space communication require high-efficiency and high-power emission and would benefit from highly efficient emission in the first atmospheric transparency window between 3--5 um. However, few QCL devices emit in this range.

Purpose

The QCL is a complex device with specifically engineered material layers that can vary in both width and material composition. For our QCL designs, we have 18 total variables: the width of the 16 layers plus barrier material composition and well material composition. With this many variables, traditional mathematical design optimization is impossible. We use an MOEA for the purpose of exploring the massive design space for optimization. However, algorithmic efficiency is incredibly important for these methods to be successful. This work explores various techniques intended to improve the search efficiency using parallelization. The indended outcome of this work is identification of which methods converge on optimized designs the quickest.

Method

An MOEA was built using simple rank and sort techniques for selection of the best candidates from generation to generation. This algorithm has parallelization, but is limited to only running the simulation of design candidates in parallel. This algorithm is used as the reference control in determining the improvement, if any, of alternate techniques. Several variations of the algorithm were built using advanced selection and parallelization techniques. For example, in these techniques the design space is divided and distributed among parallel processes. The optimization problem is solved in these more limited design spaces and results are shared among the workers to increase population diversity and reach a global optimum. Advanced candidate selection methods were also used and compared. These algorithms are being deployed on an 8-node high performance cluster to assure consistent computation capacity.

Results

Each algorithm variant is currently being validated to assure consistent evaluation of the resulting designs. Once this process is complete each algorithm will be run until reaching convergence or until a hard stop limit is reached. We will compare the run times of the parallel algorithms to the control. Testing of the control algorithm has already revealed design convergence can be achieved within 25-30 generations each with a population of 64 candidates.

Significance

Using computation methods for engineering design is a common technique, but has seen a recent expansion into highly complex multi-faceted designs. This is facilitated by the change in computing architectures to use multiple processing cores across computational nodes. While the techniques being developed in this work are targeted at the optimization of QCL devices, these techniques can be applied to nearly any multi-modal complex design problem. Furthering our understanding of how we can best utilize large computational power for efficient search is the key to making MOEAs successful when applied to complex designs.

Location

DUC Ballroom A&B

Format

Poster Presentation

Poster Session

Afternoon

This document is currently not available here.

Share

COinS
 
Apr 29th, 1:00 PM Apr 29th, 3:00 PM

Parallelization Techniques of a Multi-Objective Evolutionary Algorithm for Quantum Cascade Laser Design Optimization

DUC Ballroom A&B

Evolutionary computation is a set of optimization algorithms that use the basic processes of Darwinian nature: selection, breeding, mutation, inheritance, etc. to make decisions about a population based on objectives. This process is iterative and the intended outcome is progress, based on the objectives, in the population. In this work, the techniques of evolutionary computation, specifically multi-objective evolutionary algorithms (MOEAs), are used to search the massive design space of the diagonal transition quantum cascade laser (QCL) Quantum cascade lasers have become the standard for high-efficiency emission sources in the mid-infrared (mid-IR) and THz spectrums. Applications such as infrared counter measures, trace gas detection, and free-space communication require high-efficiency and high-power emission and would benefit from highly efficient emission in the first atmospheric transparency window between 3--5 um. However, few QCL devices emit in this range.