Generative Deep Learning Model for a Multi-Level Nano-Optic Broadband Power Splitter

Yingheng Tang, Keisuke Kojima, Toshiaki Koike-Akino, Ye Wang, Pengxiang Wu, Mohammad H. Tahersima, Devesh K. Jha, Kieran Parsons, Minghao QiOFC 2020 · DOI · OpenReview

TL;DR

Researchers have developed a novel deep learning model, a Conditional Variational Autoencoder (CVAE) with adversarial censoring, to automatically design complex nano-optic devices called power splitters. This AI-driven approach significantly accelerates the design process, producing highly efficient devices with arbitrary splitting ratios that are smaller than previously possible.

What Problem Does This Paper Address?

Designing intricate nanophotonic devices, like power splitters that divide light signals, is a computationally intensive and time-consuming process. Traditional methods, such as direct binary search or genetic algorithms, require vast numbers of electromagnetic simulations, making them impractical for exploring the immense design space. While machine learning offers a promising alternative for accelerating this process, existing deep learning approaches often struggle with generating devices that precisely meet specific performance criteria and can be difficult to train. This paper tackles the challenge of creating a more efficient, flexible, and accurate AI-driven design tool for nanophotonic devices, specifically focusing on broadband power splitters with tunable splitting ratios.

Methods

The core of this work is a Conditional Variational Autoencoder (CVAE). Unlike standard VAEs that learn a general data distribution, CVAEs incorporate "conditions" (in this case, desired performance metrics like transmission and reflection spectra) to guide the generation process. This allows the model to produce designs tailored to specific requirements.

Key innovations include:

  • Adversarial Censoring: To improve the quality and stability of the generated designs, an adversarial block was introduced. This block helps to "isolate" the latent representation of the device from its performance characteristics during training. This separation prevents the model from learning spurious correlations and leads to more robust designs.
  • Multi-level Pixel Structure: Instead of simple binary (etched or not etched) structures, the researchers use a multi-level approach where the dimensions of etched holes can vary. This adds complexity but allows for finer control over the device's optical properties.
  • Active Learning: To further enhance performance, the team employed an active learning strategy. This involves training an initial model, using it to generate a batch of designs, simulating those designs, and then using the simulation results to refine the training data for subsequent training rounds. This iterative process helps the model learn from its own predictions and improve its accuracy.

Datasets

The paper describes training data derived from binary hole vectors obtained through a direct binary search method. This initial dataset consists of 15,000 patterns. For evaluation and refinement, FDTD (Finite-Difference Time-Domain) simulations are used to verify the performance of generated devices. The active learning process involves generating 1,000 variable hole size patterns, simulating them, and then appending these labeled data points to the training set.

Key Insights and Ideas

  • CVAE for Photonics Design: Applying CVAEs with adversarial censoring is a novel approach for nanophotonic device design, demonstrating its potential to learn complex input-output relationships in this domain.
  • Adversarial Censoring Enhances Robustness: The adversarial block is crucial for disentangling device structure from performance, leading to more reliable and higher-performing generated devices.
  • Multi-level Structures Unlock Finer Control: Moving beyond binary designs to multi-level hole dimensions allows for more sophisticated optimization and better performance.
  • Active Learning Accelerates Convergence: The active learning strategy significantly boosts performance by iteratively refining the model based on simulated results, reducing the need for extensive manual design iterations.
  • Smallest Broadband Arbitrary Ratio Splitter: The generated devices are remarkably compact (2.25 x 2.25 µm²) and achieve high performance across a broad bandwidth, representing a significant advancement in miniaturization and functionality.

Key Results

  • The proposed CVAE model with adversarial censoring and active learning can generate broadband power splitters with arbitrary splitting ratios.
  • The generated devices achieve an overall performance (Figure of Merit) close to 90% across the 1250 nm to 1800 nm bandwidth.
  • The reflection is consistently below -20dB, and total transmission is greater than 87% across the specified bandwidth.
  • Compared to the preliminary model, the updated model shows a significant improvement in total transmission, for example, increasing from 85% to 90% for 1:1 splitters.
  • The average FOM for devices generated by the active learning assisted CVAE model is 0.9009, outperforming conventional CVAE and CVAE with adversarial censoring alone.
  • The devices are exceptionally small, with a footprint of 2.25 × 2.25 µm².

Limitations and Future Work

  • Training Data Dependency: While the model can generate new patterns, its performance is still dependent on the quality and diversity of the initial training data.
  • Generalization to Other Devices: The current work focuses on power splitters. Extending this framework to design other types of nanophotonic devices would be a logical next step.
  • Computational Cost of Simulation: Although the AI model accelerates design, the underlying FDTD simulations for verification remain computationally intensive. Future work could explore methods to reduce this reliance.
  • Exploration of Latent Space: Further investigation into the latent space and its relationship to device physics could lead to deeper insights and more intuitive design control.

Summary updated

Abstract

A novel Conditional Variational Autoencoder (CVAE) model with the adversarial censoring is presented to help to generate the 550nm broad bandwidth (1250nm to 1800nm) power splitter with arbitrary splitting ratio.
BibTeX
@inproceedings{tang2020generative,
  author = {Yingheng Tang and Keisuke Kojima and Toshiaki Koike-Akino and Ye Wang and Pengxiang Wu and Mohammad H. Tahersima and Devesh K. Jha and Kieran Parsons and Minghao Qi},
  title = {{Generative Deep Learning Model for a Multi-Level Nano-Optic Broadband Power Splitter}},
  booktitle = {OFC},
  year = {2020},
  doi = {10.48550/arxiv.2003.03747},
}

Loading related content...

Figure