Deep Neural Network Inverse Design of Integrated Photonic Power Splitters

Mohammad H. Tahersima, Keisuke Kojima, Toshiaki Koike‐Akino, Devesh K. Jha, Bingnan Wang, Chung-Wei Lin, Kieran ParsonsScientific 2019 · DOI

TL;DR

This paper introduces a deep neural network (DNN) approach for the inverse design of integrated photonic power splitters. By training a ResNet model on simulated data, they can rapidly design compact, high-efficiency devices with specific power splitting ratios, overcoming the computational cost of traditional optimization methods.

What Problem Does This Paper Address?

Designing nanophotonic devices, such as power splitters, involves optimizing complex subwavelength structures to achieve desired optical responses. Traditional methods like finite-difference time-domain (FDTD) simulations are computationally intensive, requiring extensive simulation time and numerous iterations for parameter sweeps. This makes the design process slow and challenging, especially for devices with a vast design space, like those with many potential structural configurations. The paper aims to accelerate this design process by leveraging deep learning for both predicting optical responses (forward design) and generating device structures for a target response (inverse design).

Methods

The core of the paper's methodology lies in using a Deep Neural Network (DNN), specifically a Residual Network (ResNet), for both forward and inverse design tasks.

  • Forward Design: The DNN takes a device's topology (represented as a 2D array of etched holes) as input and predicts its spectral transmission and reflection response.
  • Inverse Design: The DNN takes a desired spectral response (e.g., specific transmission ratios for output ports) as input and outputs the corresponding device topology.

The authors trained their ResNet model on a dataset of approximately 20,000 simulated power splitter designs. They employed a loss function based on Gaussian log-likelihood for the forward problem (regression) and Bernoulli log-likelihood for the inverse problem (classification). The ResNet architecture was chosen to overcome the vanishing gradient problem common in deep fully-connected networks, allowing for deeper architectures (up to 8 hidden layers) that improved performance.

Datasets

The primary dataset used in this study was generated through numerical simulations using Lumerical's FDTD simulation package.

  • Generated Dataset: Approximately 20,000 unique 20x20 etched hole vector (HV) configurations were simulated. Each HV was associated with its spectral transmission (T1, T2) and reflection (R) response at a wavelength range of 1450-1650 nm. This dataset was split into training (80%) and testing (20%) sets.
  • Novelty: While the paper doesn't introduce a publicly released benchmark dataset, the generated dataset represents a novel collection of simulated nanophotonic device designs and their corresponding optical responses, specifically tailored for power splitter inverse design. The dataset was curated to include both symmetric and asymmetric designs with various target splitting ratios and optimization metrics.

Key Insights and Ideas

  • DNNs as a Design Accelerator: The most significant insight is that DNNs can drastically reduce the design time for complex nanophotonic structures from hours or days (with traditional methods) to fractions of a second.
  • Inverse Design Capability: The paper demonstrates that DNNs can effectively perform inverse design, generating device geometries for specific performance targets, a capability not easily achieved with conventional forward-only simulation methods.
  • ResNet Architecture Advantage: Using a ResNet architecture proved crucial for achieving better performance by enabling deeper networks without suffering from vanishing gradients, leading to more accurate predictions and designs.
  • Compact Device Designs: The DNN successfully designed remarkably compact power splitters (2.6 x 2.6 µm²), which is significantly smaller than previously reported devices for similar functionalities.
  • Simultaneous Optimization: The approach allows for simultaneous optimization of multiple parameters, such as maximizing transmission efficiency (above 90%) while minimizing reflection (below -20 dB) and achieving target splitting ratios.

Key Results

  • Design Speed: The trained DNN can design a power splitter topology in a fraction of a second.
  • Device Performance: The designed 1x2 power splitters achieve transmission efficiencies above 90% and reflection below -20 dB at 1550 nm.
  • Device Footprint: The power splitters are remarkably compact, with a footprint of 2.6 x 2.6 µm².
  • Splitting Ratio Flexibility: The DNN can design devices for a wide range of arbitrary splitting ratios (e.g., 1:1, 1:1.05, 1:1.5, 1:2, 1:3.5).
  • Prediction Accuracy: The correlation coefficient between true numerically verified optical responses and DNN predictions was above 0.995 for both output ports.
  • Fabricability: The designs are constrained to a 20x20 vector of etched holes with a radius of 45 nm, which is considered fabricable with current semiconductor technology.

Limitations and Future Work

  • Data Generation Cost: While the design process is fast, the initial training of the DNN requires a substantial dataset generated through computationally expensive FDTD simulations. The paper notes that collecting the ~20,000 data points took about two weeks on a desktop computer.
  • Overfitting: The authors observed some overfitting in deeper ResNet configurations (e.g., 10x100 structure), suggesting that further regularization techniques might be beneficial.
  • Bandwidth: While the devices show broadband transmission (1450-1650 nm), bandwidth was not explicitly optimized as a primary objective.
  • Future Work: The authors suggest exploring dropout for regularization and increasing hidden layer depth further. They also mention the potential for hardware implementation of neural networks for even faster operation. The broader applicability of this DNN inverse design approach to other nanophotonic components is also a promising avenue.

Summary updated

Abstract

Predicting physical response of an artificially structured material is of particular interest for scientific and engineering applications. Here we use deep learning to predict optical response of artificially engineered nanophotonic devices. In addition to predicting forward approximation of transmission response for any given topology, this approach allows us to inversely approximate designs for a targeted optical response. Our Deep Neural Network (DNN) could design compact (2.6 × 2.6 μm<sup>2</sup>) silicon-on-insulator (SOI)-based 1 × 2 power splitters with various target splitting ratios in a fraction of a second. This model is trained to minimize the reflection (to smaller than ~ -20 dB) while achieving maximum transmission efficiency above 90% and target splitting specifications. This approach paves the way for rapid design of integrated photonic components relying on complex nanostructures.
BibTeX
@inproceedings{tahersima2019deep,
  author = {Mohammad H. Tahersima and Keisuke Kojima and Toshiaki Koike‐Akino and Devesh K. Jha and Bingnan Wang and Chung-Wei Lin and Kieran Parsons},
  title = {{Deep Neural Network Inverse Design of Integrated Photonic Power Splitters}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-018-37952-2},
}
Figure