Photonic design involves developing optical components and systems that use light for computing, communications, and sensing. With the growing complexity of these systems, the adoption of machine learning has proven effective in modeling intricate structures and extracting valuable insights from extensive datasets.
Image Credit: Quardia/Shutterstock.com
Machine Learning and Photonics: A Historical Background
In the 1990s, artificial neural networks (ANNs), a branch of machine learning, made its way into photonics research, initially in microwave design. Using ANNs as a computer-aided design tool proved successful for designing transmission lines, vias, filters, amplifiers, and antennas.
The interplay between microwave research and ANNs was not coincidental, as microwave device designs often involve optimizing parameters for a given target, aligning well with the capabilities of ANNs. However, more complex photonic designs faced limitations until the recent advancements in machine learning.
The past decade witnessed the rise of machine learning, marked by innovations like ReLU activation, dropout, and batch normalization, enabling the design and training of deeper neural networks with larger datasets and better performance.
Machine Learning Algorithms Used in Photonic Design
Machine learning algorithms in photonic design are categorized into supervised and unsupervised learning.
Supervised learning uses labeled data for training, optimizing algorithms like convolutional neural networks (CNNs) to correlate photonic structures with optical properties and achieve specific optical functions. On the other hand, unsupervised learning with unlabeled data determines novel patterns, emphasizing feature extraction and eliminating the need for extensive labeled datasets in predicting optical responses.
Convolutional Neural Networks (CNNs)
CNNs leverage built-in spatial processing via convolution layers to extract features from simulation and experimental images automatically. This makes them well-suited for designing and optimizing parameters of photonic crystals and metamaterials based on optical microscopy data. They have been employed to tweak the positions of elements like holes in nanocavities to achieve quality factors exceeding one billion.
Multilayer Perceptron (MLP)
MLPs composed of multiple hidden layers are widely used to map complex relationships between optical properties and photonic structures. They can be trained on scattering spectra and other optical data to predict nanostructures that produce target responses. Their key advantages include flexibility in learning nonlinear interactions and restricting outputs to feasible photonic designs.
Recurrent Neural Networks (RNNs)
RNN architectures incorporating cyclical connections suit modeling continuous optical signals in the time domain. Long short-term variants like LSTMs can overcome issues like vanishing gradients. This facilitates using them as generative models to simulate realistic sequential optical responses, which helps design dynamic photonic systems.
Generative Adversarial Networks (GANs)
GANs consist of generator and discriminator models engaged in adversarial training. The generator learns to produce increasingly realistic fake photonic designs while the discriminator gets better at distinguishing fakes. Once trained, the generator efficiently produces diverse, high-quality structures matching complex target optical properties.
How is Machine Learning Used in Photonic Design?
Optical Nanoparticle Design
Core-shell nanoparticles enable exotic optical phenomena relevant to imaging and spectroscopy applications. However, their complex geometries make conventional design challenging. Consequently, researchers are applying machine learning models like multilayer perceptrons (MLPs) to link structural parameters to optical responses.
In 2018, Peurifoy et al. demonstrated using MLPs for both forward modeling and inverse design of multilayered silicon/titanium dioxide nanoparticles. The MLPs learned to accurately predict scattering spectra and output corresponding layer configurations on demand by training on computationally generated datasets. This avoided repetitive simulations traditionally needed to determine suitable structures.
So et al. (2019) extended this approach by integrating classification and regression techniques in the inverse design of optical material and structural thickness. This dual-task strategy simultaneously identified materials in each layer (via classification) and predicted thickness (via regression), effectively addressing the challenges of inverse design.
This approach allowed for comprehensive and precise optimization of optical properties in designed structures, showcasing the versatility of machine learning in concurrent photonic property and structural optimization.
Metasurface Design
Metasurfaces impart exceptional wavefront and polarization control unachievable with conventional optics. However, the design complexity scales exponentially with the number of nanoscale constituent elements. Therefore, researchers are applying generative deep learning to construct large-scale metasurfaces efficiently.
A study published in Advanced Materials combined generative adversarial networks with an evolutionary algorithm for configurable optical metasurface design. This produced a specialized network for generating metamolecule patterns to manipulate light with unprecedented dexterity.
Such divide-and-conquer approaches enabled by machine learning are indispensable for tailoring the responses of extensive metasurfaces.
Integrated Photonics
Integrated photonic circuits miniaturize optical processing onto chips, but the design intricacy grows with greater on-chip functionality. Consequently, machine learning aids tasks like multi-objective topology optimization and fabrication resilience.
In 2022, Mao et al. presented a multi-task deep neural network for joint optimization of key integrated photonics components. By inputting only low-resolution maps of photonic structures, their model could simultaneously enhance performance metrics like transmission efficiency, bandwidth, and mode overlap.
Separately, Yesilyurt et al. (2023) introduced a neural network-based inverse design technique, incorporating differentiable optical solvers to improve fabrication tolerance (by accounting for simulated systematic and random nonidealities), offering a practical approach for designing high-performance single-material multilayer film stacks across diverse applications.
Overall, the continuity afforded by deep neural networks suits the extensive configuration spaces and fabrication uncertainties associated with integrated platforms well.
Image Reconstruction
Inverse problems in computational imaging aim to reconstruct images from incomplete information, like fewer pixel measurements. Recently, CNNs have become ubiquitous for such tasks, leveraging their pattern completion abilities.
A study published in Science Advances employed a deep encoder-decoder network for lensless imaging, reconstructing high-fidelity images from raw photon counts on a limited number of multiplexed pixel detectors.
Leveraging deep learning’s sample complexity reduction, this approach achieved lensless reconstruction of a 3D image with a high diffraction efficiency of 78% and an ultra-wide viewing angle of 94%, showcasing significant potential in applications such as multiplexed displays and encryption.
Powering a New Era of Photonic Devices
The unique challenges of photonic design make it fertile ground for transformative machine-learning approaches.
Future progress in this field could be guided by standardized datasets and models for transfer learning to enhance accessibility. Developing physical interpretability in machine learning models may unveil new physics insights while expanding design frameworks to cover a broader range of optical effects could unlock novel device functionalities. Additionally, leveraging algorithms like recurrent neural networks for dynamical photonics systems may lead to advancements such as self-tuning lasers.
More from AZoOptics: Methodologies for Non-Destructive Testing of Paint Layers
References and Further Reading
Ma, W., Liu, Z., Kudyshev, Z. A., Boltasseva, A., Cai, W., & Liu, Y. (2021). Deep learning for the design of photonic structures. Nature Photonics, 15(2), 77-90. https://doi.org/10.1038/s41566-020-0685-y
Duan, B., Wu, B., Chen, J. H., Chen, H., & Yang, D. Q. (2022). Deep learning for photonic design and analysis: Principles and applications. Frontiers in Materials, 8, 592. https://doi.org/10.3389/fmats.2021.791296
Peurifoy, J., Shen, Y., Jing, L., Yang, Y., Cano-Renteria, F., DeLacy, B. G., … & Soljačić, M. (2018). Nanophotonic particle simulation and inverse design using artificial neural networks. Science advances, 4(6), eaar4206. https://doi.org/10.1126/sciadv.aar4206
So, S., Mun, J., & Rho, J. (2019). Simultaneous inverse design of materials and structures via deep learning: demonstration of dipole resonance engineering using core–shell nanoparticles. ACS applied materials & interfaces, 11(27), 24264-24268. https://doi.org/10.1021/acsami.9b05857
Liu, Z., Zhu, D., Lee, K. T., Kim, A. S., Raju, L., & Cai, W. (2020). Compounding meta‐atoms into metamolecules with hybrid artificial intelligence techniques. Advanced Materials, 32(6), 1904790. https://doi.org/10.1002/adma.201904790
Mao, S., Cheng, L., Chen, H., Liu, X., Geng, Z., Li, Q., & Fu, H. (2022). Multi-task topology optimization of photonic devices in low-dimensional Fourier domain via deep learning. Nanophotonics, 12(5), 1007-1018. https://doi.org/10.1515/nanoph-2022-0361
Yesilyurt, O., Peana, S., Mkhitaryan, V., Pagadala, K., Shalaev, V. M., Kildishev, A. V., & Boltasseva, A. (2023). Fabrication-conscious neural network based inverse design of single-material variable-index multilayer films. Nanophotonics, 12(5), 993-1006. https://doi.org/10.1515/nanoph-2022-0537
Ren, H., Shao, W., Li, Y., Salim, F., & Gu, M. (2020). Three-dimensional vectorial holography based on machine learning inverse design. Science advances, 6(16), eaaz4261. https://doi.org/10.1126/sciadv.aaz4261
Brunner, D., Soriano, M. C., & Fan, S. (2023). Neural network learning with photonics and for photonic circuit design. Nanophotonics, 12(5), 773-775. https://doi.org/10.1515/nanoph-2023-0123
Hamerly, R. (2021). The future of deep learning is photonic. [Online]. IEEE Spectrum. Available at: https://spectrum.ieee.org/the-future-of-deep-learning-is-photoni
