[Paper Review] Deep Neural Network Inverse Design of Integrated Nanophotonic Devices

Mohammad H. Tahersima, Keisuke Kojima+5 more|arXiv (Cornell University)|Sep 10, 2018

Photonic and Optical DevicesEngineering35 references17 citations

TL;DR

This paper proposes a deep neural network (DNN)-based inverse design framework for integrated nanophotonic devices, enabling rapid, sub-second generation of compact silicon-on-insulator (SOI) 1×2 power splitters with user-defined splitting ratios. The method achieves >90% transmission efficiency and reflection <−20 dB across broadband (1450–1650 nm) operation, demonstrating the first DNN approach to simultaneously optimize for high efficiency, low reflection, and arbitrary splitting ratios in a single inverse design framework.

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 (2.6x2.6 μm2) silicon-on-insulator (SOI)-based 1 X 2 power splitters with various target splitting ratios in a fraction of a second. This model is trained to minimize the reflection (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.

Motivation & Objective

• To address the computational bottleneck of optimizing complex nanophotonic devices using traditional FDTD simulations, which are time-consuming and impractical for large design spaces.
• To develop a deep learning framework capable of both forward and inverse modeling of nanophotonic device responses, enabling rapid design by specification.
• To simultaneously optimize for high transmission efficiency (>90%), low reflection (<−20 dB), and arbitrary splitting ratios in a single inverse design process.
• To demonstrate that a DNN can generalize across a vast design space (2^400 combinations) using only ~20,000 FDTD simulations, reducing design time from hours to milliseconds.

Proposed method

• A residual neural network (ResNet) architecture is trained to map binary hole position matrices (device topology) to spectral transmission and reflection responses (forward model).
• An inverse design model is trained to map desired spectral responses (target splitting ratios and low reflection) to corresponding binary hole position vectors (device topology).
• The inverse model uses a Bernoulli distribution classifier to predict hole positions (0 or 1), with quantization applied post-prediction to generate binary designs.
• Training data is generated via 3D FDTD simulations using Lumerical’s FDTD package, with dispersive refractive indices for Si and SiO2 over 1.45–1.65 µm.
• The network is trained using TensorFlow in Python, with loss minimized over multiple epochs to ensure convergence to binary solutions (0 or 1).
• Validated designs are re-simulated via independent FDTD to confirm performance, including electric field propagation at 1550 nm.

Experimental results

Research questions

• RQ1 Can a deep neural network accurately predict the optical response of a nanophotonic device from its geometric topology?
• RQ2 Can a DNN perform inverse design to generate a device topology that achieves a user-specified optical response, including arbitrary splitting ratios and low reflection?
• RQ3 Can a DNN generalize across a vast design space (2^400 combinations) with only ~20,000 training simulations?
• RQ4 Can the inverse design framework simultaneously achieve high transmission (>90%), low reflection (<−20 dB), and broadband operation (1450–1650 nm)?

Key findings

• The DNN achieved transmission efficiency exceeding 90% for all tested splitting ratios (1:1, 1:2, 1:3), with the highest reported efficiency in integrated power splitters to date.
• Reflection was reduced below −20 dB for all cases except the 1:2 splitter, demonstrating effective suppression of back-reflection in the design.
• The inverse design process generated valid, optimized device topologies in less than a second per design, enabling rapid, interactive design by specification.
• The method achieved broadband operation across 1450–1650 nm, with consistent performance across all target splitting ratios, indicating robustness to wavelength variation.
• The trained DNN generalized across the entire design space (2^400 combinations) using only ~20,000 FDTD simulations, significantly reducing the need for exhaustive numerical optimization.