SUMMARY

Periodic gratings utilized as emitters increase the efficiency of thermophoto-
voltaic (TPV) systems. These gratings work by altering the emittance spectrum
incident on the photovoltaic cell to better match the band gap of the cell. Photons
at slightly higher energies than the band gap are the most efficient as they generate
electron hole pairs while minimizing thermalization losses. This prompts the use of
gratings to be used as selective emitters. Even for a one-dimensional (1D) grating,
millions of possible geometries exist and simulating even a fraction is infeasible. This
prompts the use of metaheuristics. It should be noted that due to the stochastic
nature of these optimization methods, a globally optimal solution is not guaranteed,
and instead these methods seek to provide “close enough” solutions.
Generally, metaheuristic algorithms have been extensively studied and compared
with each other; according to the “no free lunch” (NFL) theorem, all optimization
algorithms are equivalent when averaged over all possible problems. Therefore, a com-
parison of existing algorithms for the optimization of a system, composed of a 2,000
K 1D tungsten grating paired with a 300 K InGaSb cell, was performed. After using
the comparison, a hyperheuristic optimization was used to algorithmically develop
a purpose-built metaheuristic algorithm. Rigorous coupled wave analyses (RCWA)
take too long to natively perform for the hyperheuristic search. Fully connected
neural nets (FCNN) solve this problem when used as surrogate models. The new
optimization algorithm created in this way showed significantly better performance
than all the existing algorithms it was compared against. Then, this algorithm was
used to optimize emitters for a normalized emittance spectrum, maximum efficiency,
and maximum power.