Fabrication-Aware Inverse Design | Tidy3D Learning Center

Fabrication-Aware Inverse Design

The October 9, 2025 seminar walks through a complete dual-layer grating coupler workflow: start from a uniform baseline, pull a strong seed design with Bayesian optimization, switch to adjoint gradients for per-tooth control, study fabrication sensitivities, and close the loop with measurement-driven calibration. Everything runs inside Tidy3D, so you can rerun the exact same jobs or adapt the utilities to your own device stack.

Archived seminar material. These saved outputs were produced with Tidy3D 2.9.1

Notebook Lineup

Setup Guide: Building the Simulation

Builds the nominal SiN stack, launches the reference simulation, and visualizes the initial geometry so the later notebooks can reuse the cached job ID.

Bayesian Optimization: Finding a Strong Baseline

Uses a five-parameter Bayesian search to quickly find a good uniform grating. This provides a practical baseline before investing in gradients.

Adjoint Optimization: High-Dimensional Refinement

Expands to per-tooth parameters and applies Adam with adjoint sensitivities to apodize the grating and boost efficiency.

Fabrication Sensitivity Analysis: Is Our Design Robust?

Sweeps ±20 nm etch bias, runs Monte Carlo samples, and logs adjoint-derived sensitivity units (Δ objective / Δ parameter) so readers understand what the gradients mean physically.

Robust Adjoint Optimization for Manufacturability

Penalizes variance across nominal/over/under corners, illustrating a fabrication-aware adjoint loop that matches what we demoed live.

Monte Carlo View: Nominal vs Robust Grating

Reruns the Monte Carlo campaign for both nominal and robust devices to quantify yield improvements.

Measurement Calibration: Bridging Simulation and Fabrication

Demonstrates gradient-based calibration of tooth widths against (synthetic) spectra, using adjoint sensitivities to recover the as-fabricated geometry from optical measurements.

Getting the Code

Use the .ipynb download button on each notebook page. You will also need the helper scripts setup.py and optim.py plus the result checkpoints linked below.

How to Run the Series

Install tidy3d and bayesian-optimization (pip install tidy3d bayesian-optimization) and configure your API key.
Execute the notebooks in order; each step writes results into results/ and later notebooks assume those JSON files exist.

Supporting Assets

setup.py — shared simulation builders, fabrication constraints, and helper functions.
optim.py — a lightweight, autograd-friendly Adam implementation with parameter clipping.
results/ — Bayesian seed, adjoint design, and robust design checkpoints consumed by subsequent notebooks.