Matlab 2015b or newer. Older versions may be sufficient but have not been tested.
RETICOLO - rigourous coupled wave analysis (RCWA) solver. Can be downloaded from RETICOLO. Copy the folder reticolo_allege into the working directory.
Run RunOpt.m with default parameters. The example optimization should begin immediately if all files have been installed corrected.
In RunOpt.m, define all optimization parameters as necessary. Descriptions of all parameters can be found in Functions/Initialize.m along with their default values.
A schematic of metagrating parameter defintions can be found at MetaNet.
Robustness parameters, found in OptParm.Optimization.Robustness, are accepted as vectors of dynamic length according to the number of robustness simulations used in computing the gradient. Details on robust optimization can be found in this paper.
The default robustness parameters of
StartDeviation = [-5 0 5];
Weights = [.5 1 .5];
define a gradient derived from a -5nm eroded structure, an unperturbed structure, and a 5nm dilated structure, weighted at 0.5x, 1x, and 0.5x respectively.
Additionally, the magnitude of robustness can be scaled as the optimization progresses between the values defined in StartDeviation and EndDeviation. The speed of scaling is defined by Ramp. For no scaling, StartDeviation and EndDeviation should be set to the same value.
Additional robustness can be added as desired. i.e.
StartDeviation = [-10 -5 0 5 10];
Weights = [.25 .5 1 .5 .25];
defines additional simulations with even greater perturbations.
Optimizations without robustness can be specified by
StartDeviation = [0];
Weights = [1];
The optimization method can be parallelized across multiple threads in the case of either two polarization devices or robust optimization. In OptimizeDevice.m, the for loops
for robustIter = 1:NRobustness
and
for polIter = 1:NumPol
can be replaced with parfor loops instead. Note that only one of the two loops may be parallelized at a time.
Checkpointing saves the state of the optimization and allows the user to stop and start the optimization without losing progress. This behavior is desirable when running shared computing resources where jobs may be preempted without warning.
To enable checkpointing, set OptParm.Checkpoint.Enable = true and define a checkpoint file location with OptParm.Checkpoint.File. Then, throughout the optimization, the entire optimization state will be saved every few iterations with a frequency defined by OptParm.Checkpoint.Frequency.
When restarting a checkpointed optimization, the optimization will attempt to load the file at OptParm.Checkpoint.File if it exists, and continue, otherwise the optimization will start from the beginning again.
Warning: Checkpointing may fail if the optimization is terminated while writing to the checkpoint file.
Please cite this code as:
MetaNet: A new paradigm for data sharing in photonics research
Jiaqi Jiang, Robert Lupoiu, Evan W. Wang, David Sell, Jean Paul Hugonin, Philippe Lalanne, Jonathan A. Fan