Inverse Design of Photonic Integrated Devices: Neural Network Tutorial

This repository provides Python code accompanying the tutorial paper on inverse photonic design using neural networks, as described in the associated manuscript.

---

Overview

This tutorial is structured into clear steps corresponding to the sections of the paper, covering:

• Forward Model: Training a neural network model to predict TE coupling coefficients from geometry parameters.

  • Simple forward model (N_MODELS = 1, AUGMENT_DATA = False)
  • Simple forward model with data augmentation (N_MODELS = 1, AUGMENT_DATA = True)
  • Ensemble forward model (N_MODELS > 1, AUGMENT_DATA = True recommended)

• Inverse Model: Predicting geometry parameters from specified TE coupling coefficients.

  • Simple inverse neural network (without tandem)
  • Tandem inverse neural network (with pre-trained forward network)

The provided Python script allows you to reproduce these examples directly, using simple flags and parameters. Additionally, the folder ./Simulations_setup contains example code and needed scripts to run Ansys Lumerical simulations for data generation.

---

Getting Started

Step 1: Clone the Repository

git clone https://github.com/maldig/inverse-photonics-design.git
cd inverse-photonics-design

Step 2: Environment Setup

Create and activate a Python virtual environment (recommended):

python -m venv env
source env/bin/activate   # Linux/Mac
.\env\Scripts\activate    # Windows

Install dependencies:

pip install -r requirements.txt

---

Data Files

Ensure the following CSV files are present in the root folder:

• Training data:

  • data_v19.csv (original)
  • data_v19_aug.csv (augmented)

• Test data:

  • frontier_v19.csv

• Generation data (optional):

  • filtered_points_3.csv

---

Running the Code: Neural Network Tutorial Sequence

Follow these instructions sequentially to reproduce results from the manuscript.

Section 1: Training the Forward Model

1a. Simple Forward Model

Edit flags in inverse_design.py:

TRAIN_FORWARD = True
N_MODELS = 1            # Simple model
AUGMENT_DATA = False    # Original data only

Then run:

python inverse_design.py

1b. Simple Forward Model with Data Augmentation

Edit flags:

TRAIN_FORWARD = True
N_MODELS = 1
AUGMENT_DATA = True     # Augmented dataset

Then run:

python inverse_design.py

1c. Ensemble Forward Model

Edit flags:

TRAIN_FORWARD = True
N_MODELS = 5            # or any number >1
AUGMENT_DATA = True     # Augmented dataset recommended

Then run again:

python inverse_design.py

---

Section 2: Training the Inverse Model

2a. Simple Inverse Neural Network (illustrative)

Swap X and y manually in inverse_design.py:

X_train_scaled, y_train_scaled = y_train_scaled, X_train_scaled
X_val_scaled, y_val_scaled = y_val_scaled, X_val_scaled
X_test_scaled, y_test_scaled = y_test_scaled, X_test_scaled

Edit flags:

TRAIN_FORWARD = False
TRAIN_INVERSE = True
N_MODELS = 1

Run:

python inverse_design.py

2b. Advanced Tandem Inverse Neural Network (Recommended)

Ensure forward model is trained first, then edit flags:

TRAIN_FORWARD = False
TRAIN_INVERSE = True
N_MODELS = 5

Run:

python inverse_design.py

---

Section 3: Generating Designs with the Trained Inverse Model

Set flags in the script:

TRAIN_FORWARD = False
TRAIN_INVERSE = False
TEST_MODE = False

Place TE targets in filtered_points_3.csv. Run:

python inverse_design.py

Predicted design parameters saved:

results/generated_values_tandem_v19.csv

---

Directory Structure after Execution

inverse-photonics-design/
├── data/                         # Optional
├── data_v19.csv
├── data_v19_aug.csv
├── frontier_v19.csv
├── filtered_points_3.csv
├── models/
│   ├── model_1.pt
│   ├── ...
│   ├── pytorch_ensemble_model_v19.pt
│   └── inverse_model_v19.pt
├── results/
│   └── generated_values_tandem_v19.csv
├── inverse_design.py
├── requirements.txt
└── README.md

---

Customization

• To modify neural network architectures, edit ForwardNet and InverseNet classes.
• Experiment with training parameters (batch size, epochs, learning rate) at the top of inverse_design.py.

---

Reference

---