In this project we implement the optimization algorithm in ref [1] on an optoelectronic Ising machine that we built similar to the one in ref [2]. The difference is that we implement on a Coherent Ising Machine (CIM) as opposed to D-Wave quantum annealers which were used as the Ising solver in [1]. Using the CIM is advantageous since all-to-all connections can be implemented with it whereas in D-Wave computer only sparce connections are available. The optimization problem that we chose is 2D optimization of photonics devices.
The followin is the experimental setup for the CIM at IIT Madras.
The following image illustrates the algorithm in [1]. The changes in our project is we use a CIM instead of the D-Wave anealer. We are optimizing a size constrained Silicon Photonic 3dB directional coulpler.
The inverse designed dataset is generated from several python based FDTD/FDFD packaged such as MEEP, ceviche, Angler etc. One devide of interest is a silicon photonic 3dB directional coupler, but also datset of defractive metastructures can be optained from Metanet as well. MEEP was used for final design verification but ceviche and Anglere were used for fast generation of datasets.
In case of generating dataset using MEEP, a computing cluster such as AWS can be used.(to do)
The bVAE used here is structured as: x input layers -> hidden layer(512 nodes)-> hidden layer (xx)-> bottleneck layer(500??)-> gumbell layer (500??)-> hidden layer(xx)-> hidden layer(512)->output layer(x)