Designing GaAs Solar Cells Using Genetic Algorithms
 

by
G.B. Lush, Richard Zamudio, and Steven Hernandez
August 1999



We have completed a tool that runs simulations of GaAs solar cells using device parameters provided by a Genetic Algorithm.

We hope in the future to develop a tool that will allow us to distribute the processing load by running many simulations across our Sun workstation Unix network. We are calling this a Simulation Manager because it will distribute computing jobs to all the workstations, according to their ability to compute. It will keep track of how quickly each computer returns jobs, and send the next jobs to the fastest processors.

The plots you will see below are interesting because they illustrate how a GA searches and finds solutions to the questions we pose. The plots are of the parameters we have as design parameters for a GaAs solar cell. What you will see as you look at these plots is that the parameters that are most important in the design of a solar cell--such as mole fraction and thickness of the window layer--converge to the best value right away, and the parameters that have less impact on the solution--such as window layer doping--take longer to converge.

The first plot is of the mole fractions of the window and of the rest of the solar cell. Note that the mole fraction and thickness of the window (thickness on next graph), which have the strongest effect on the device efficiency, find themselves at the ideal values after just a few generations. Solar cell mole fraction stabilizes after about 25 generations.

Plot 3 shows that the base doping, certainly the most important of the dopings due to its influence on the diffusion length in the base, converges after 25 generations. The other dopings take longer. The base thickness converges after about 50 iterations.

The behaviors of the emitter doping and of the emitter thickness are interesting. Note that the emitter thickness slowly decreases over the generations, and that once it reaches a certain point, the emitter doping can increase. Higher emitter doping yields a higher Voc as long as the diffusion length is still long enough to collect the carriers generated in the emitter.

Remember that if the ideal value of a parameter is the minimum or the maximum value allowed, it can take longer to find that because it requires a mutation or randomly-chosen value to choose the minimum value or for an offspring to be created from parents with just the right values and just the right randomly-chosen beta factor for interpolation/extrapolation.