The latest 2019b R2 release comes with support for topology optimization. In this example we demonstrate how to design a 50:50 Y-splitter using the new feature. This example complements the existing Y branch Inverse design example
To use this feature, you require a license for the “Automation API”. If you have a valid license and still encounter a licensing error, please contact your Lumerical sales contact to obtain an updated license.
Introduction to Topology Optimization
Inverse design using Topology Optimization enables automatic generation of devices that are optimized for a user-specified figure of merit (FOM). In this Y-splitter example, we want to maximize T_up, the fraction of the total input energy that is transmitted into the upper arm. Since we will enforce symmetry in y-direction, the maximal value of the FOM will be 0.5. For visualization, instead of directly plotting T_up, we show the insertion loss, so FOM=0.5 - T_up.
The important difference from shape optimization is that for topology optimization, you only need to specify an optimization area. There is no need to parameterize the geometry. The optimization process will often result in non-intuitive but highly efficient structures.
The main challenge with structures generated via topology optimization is manufacturability. To control the smallest features, the optimizer continuously smooths the structure using a filter with a user-defined radius filter_R. This removes small features and rounds sharp corners and thereby improves the manufacturability.
Step 1: Define base simulation project
First, we need to define a base simulation project which contains the general structure of the device. This includes the simulation region with the input and output waveguides as well as the source and monitor positions. In this example, we optimize a 50:50 splitter over the wavelength range from 1450nm to 1650nm. The base simulation project is generated via the script file splitter_base_2D_TE_topology.lsf. We can inspect the setup by opening and running the script in FDTD:
In this example, the setup script also generates a structure group called ‘initial_guess’ which provides an initial device which can be used as a starting point for the optimization. This structure group is optional and topology optimization will also work without it .
Step 2: Define optimization region
For the next step, we first need to define several optimization parameters:
|Name||Value in the example||Description|
|size_x||3000 [nm]||Length of the device|
|size_y||1800 [nm]||Since we use symmetry along the y-axis, the value is half the height of the device|
|delta_x||20 [nm]||Size of a pixel along x-axis|
|delta_y||20 [nm]||Size of a pixel along y-axis|
|filter_R||500 [nm]||Radius of the smoothing filter which removes small features and sharp corners|
|eps_max||2.8^2=7.84||Effective permittivity for a Silicon waveguide with a thickness of 220nm|
|eps_min||1.44^2=2.0736||Permittivity of the SiO2 cladding|
In addition, we need to specify the FOM and additional optimizer settings. This is done in the script file splitter_opt_2D_TE_topology.py. Here, we also define the initial conditions. As discussed above, the main advantage of topology optimization is that there is no need for a sophisticated starting condition. Instead we can simply fill the entire optimization region with either eps_max or eps_min.
Step 3: Run topology optimization
Running the script splitter_opt_2D_TE_topology.py will start the optimization. Topology optimization as implemented here works in two phases. First, during the ‘grayscale phase’, the optimizer is free to use any permittivity value ϵ ∈ [eps_low,eps_high ] for the permittivity. Once the optimizer has reached an optimum, it will start the ‘binarization phase’. Here, the optimizer now tries to convert the continuous permittivity distribution into a binary structure which only uses either eps_low or eps_high, but no intermediate values.
The final plot after around 500 iterations should look like this :
In the above plot, we can observe spikes where the FOM temporarily gets worse. Each spike signifies an increase in the binarization factor with a subsequent optimization to improve the FOM again. Towards the end of the optimization we observe a slight overall deterioration of the FOM. This is again caused by the fact that the binarized structures performs slightly worse than a smooth structure. However, the FOM is plotted on a semi-logarithmic scale and the absolute deterioration is relatively small.