Integrated Photonic Neural Networks with Arbitrary Capabilities

Project Information

NeuroPhotonics

Grant agreement ID: 101032147

DOI

10.3030/101032147

Project closed

EC signature date 15 March 2021

Start date 1 October 2021

End date 30 September 2023

Funded under

• EXCELLENT SCIENCE - Marie Skłodowska-Curie Actions

Total cost

€ 157 355,52

EU contribution

€ 157 355,52

Coordinated by

KOC UNIVERSITY
Türkiye

Periodic Reporting for period 1 - NeuroPhotonics (Integrated Photonic Neural Networks with Arbitrary Capabilities)

Reporting period: 2021-10-01 to 2023-09-30

Summary of the context and overall objectives of the project

Continued economic and societal growth depend on complex, multi-functional, and advanced circuitry for accelerating connection to large volumes of data. Photonic integrated circuits represent a critical class of devices that enable this data access and processing capabilities. So far, photonic circuits have been designed using either traditional methods, or machine-based optimizers that are programmed to determine device geometries from scratch. While these approaches are viable for certain basic categories of photonic functions, more complex and custom applications require development of more advanced and high-performance photonic design methods in order to meet the needs of these future applications, and to standardize the photonic design ecosystem across multiple different industries. To this end, the overall goal of the NeuroPhotonics project is to build an experimentally verified photonic neural network device architecture and framework, for the design of photonic integrated circuits with arbitrary on-chip capabilities. Using arrays of customized photonic interferometers, this framework enables previously elusive optical functionality including arbitrary combinations of ultra-wideband, fabrication-tolerant, and wavelength-selective optical responses. The overall objectives include creating an open-source software package for rapidly extracting guided wave parameters, building a multi-purpose design framework for universal photonic neural network architectures for desired arbitrary functionality, experimentally demonstrating input-output mapping capabilities by fabricating and characterizing these photonic networks, and bridging project innovations with industry through a post-project innovation management plan for commercialization of the framework. This presented design platform provides a tractable path towards the systematic design of large-scale photonic systems with custom and broadband power, phase, and dispersion profiles for use in multi-band optical applications including high-throughput communications, quantum information processing, and medical/biological sensing.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

The work performed since the beginning of the project includes several computational and experimental tasks. First, an open-source silicon photonics toolkit (“SiPkit”) software library was created in order to rapidly return guided wave parameters in silicon waveguides. This package was equipped with automatic differentiation capabilities to enable gradient-based device optimization, achieve compatibility with the further stages of the project, as well as general applicability through other existing machine learning-based design tools. Following this package, a design framework for arbitrary photonic networks was created. This framework enables user-defined optical transfer functions to be realized by designing task-specific device geometries in photonic networks. The networks in this project are constructed from an input stage, cascaded and repeated custom interferometric layers, and the output stage. Each interferometric layer consists of Mach-Zehnder interferometers and individually-optimized waveguide taper structures. These custom waveguide tapers are instrumental in achieving specific phase/dispersion profiles that are automatically determined according to the specified target functionality of the device. Then, an matching artificial neural model is created according to the network topology of the photonic device, and used in calculating the transfer function for each input-output pair. This transfer function is compared with the target transfer function; and an error is calculated from their difference. This error is minimized by modifying the geometrical parameters of the photonic network, in order to optimize the device design according to the target functionality specified.

Using this capability, proof-of-concept networks have been fabricated using EUROPRACTICE foundries, experimentally characterized. All devices have shown state-of-the-art performance metrics including low insertion losses and widest bandwidths demonstrated for their respective device classes. These results have been disseminated and communicated through multiple different channels throughout the project. The entire design framework with simulation-based demonstrations of arbitrary transfer functions together with the experimental results have been demonstrated in a high-impact publication that has recently been accepted for publication. The open-source software package SiPkit has been published with an accompanying user guide. A total of seven conference/workshop presentations have been given/scheduled targeting individual results of the project. Results have been shared with academic and industrial partners through a recent workshop. A post-project innovation management plan has also been created with input from industrial users for exploitation and potential commercialization of the design framework.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

The demonstrated devices and the design framework present a novel approach to the creation of photonic circuits that are able achieve arbitrary optical functions. This overall objective has been shown through multiple independent device simulations, and also through experimental characterizations of the fabricated proof-of-concept devices. Instead of modifying several waveguide parameters by hand, the demonstrated design capabilities achieve a much larger class of previously elusive photonic functionalities including both wavelength-specific and broadband optical responses, inherent capabilities for maintaining device performance under fabrication variations, scalability to complex target functionalities such as randomly-distributed output power mapping to an arbitrary number of outputs, and multi-objective capabilities showing single devices performing multiple different functions for different input-output pairs. These capabilities represent not only the state-of-the-art performance in typical metrics including insertion loss and operation bandwidth, but also illustrate how these photonic networks enable capabilities that are otherwise impossible or computationally infeasible to implement using other design techniques.

NeuroPhotonics has implemented multiple impact streams to target academic, societal, industrial, and commercial contributions. Resulting from the project outputs, an open-source software library, an accompanying user guide, and a high-impact journal publication has been achieved. Seven conference/workshop talks have been given, including real-time demos illustrating different use cases and capabilities of the design framework. From an applications perspective, the different demonstrations with a variety of photonic functionalities including specific power, phase, and dispersion profiles have been shown in simulations and in experiments. These results are directly used in forming an innovation management roadmap with plans to incorporate the framework as a part of photonic design ecosystems, as well as using the resulting devices in product development kits of foundries. With this roadmap, meetings with industrial partners have been completed, and will continue to drive use cases of the presented design framework for applications in optical communications, sensing, and computing.

Photonic network architecture and components

Optimization of an example 1-input 4-output photonic network