Inverse design

kind: "optimization", model lightsmith.inverse.OptimizationProblem. Density-based topology optimisation of one or more design regions inside a device, driven by adjoint gradients.

Fields

Field	Meaning
device	The DeviceProblem being optimised. Its ports and fixed structures stay as given.
design_regions	DesignRegion entries. Each has a box (center_nm, size_nm), material_low and material_high, an initial density (init), filter_radius_nm, eta, an optional min_feature_nm, and optional mirror symmetry (x, y).
objective	A sum of terms. SParameterTerm is a port-to-port power or amplitude. FieldIntensityTerm is the mean |field|² in a box. Each term has a goal (maximize, minimize or target), a weight, and an aggregate over wavelengths (mean, or a smooth min or max).
optimizer	OptimizerSettings: method (mma, adam or lbfgsb), max_iter, β-continuation (beta_start, beta_max, beta_growth, beta_every), and the length-scale constraint weights.
backend	The differentiable device solver. Default lightsmith-fdfd.
options	Overrides for the device solver's options

Result

OptimizationResult contains:

• history;
• objective_initial, objective_final, and objective_binary (the score of the thresholded design);
• designs, the density maps;
• structures, the exported polygons as prisms;
• final_device;
• sparameters, from re-solving the exported design;
• provenance and warnings.

Export the design with lightsmith.inverse.export.write_gds, or with to_gdsfactory when the layout extra is installed.

Running

• ls.solve(problem) runs the optimisation and validates the final design.
• lightsmith.inverse.optimize(problem, device=..., callback=..., checkpoint="run.npz") adds progress callbacks and resumable checkpoints.
• Both need the gpu extra (PyTorch), even on a CPU.

See Inverse design for the method, and examples/inverse_design_bend.py for a complete 90° silicon bend.