INFORMATION TECHNOLOGY

“Knowledge Inheritance Network”: Design and Assembly of Distributed Intelligent Metasurfaces


The interaction of machine learning and optics/photonics is changing the way we design novel metasurface structures and develop intelligent metasurface devices. In metasurface work, a key step is to obtain a high-performance metasurface structure or distribution with the desired optical response.

Compared with the traditional electromagnetic numerical calculation methods, the metasurface reverse design based on machine learning greatly improves the design efficiency and shortens the solution time. However, as a data-driven algorithm, data is as important to machine learning as the “fuel” of the engine, and the appropriate algorithm model is even more related to the quality of the engine’s “engine”.

One major hurdle in existing explorations is that both datasets and networks are disposable. For each new state or task, all datasets and networks must be discarded and rebuilt, resulting in huge wasted resources and time consumption. In existing metasurface reverse design, each metasurface task is physically unrelated and data utilization is extremely inefficient. Therefore, if a network corresponding to the physical connection between metasurfaces can be constructed, a large number of metasurface design problems can be robustly handled, opening up a new path for the design of intelligent superdevices.

Recently, the team of Professor Chen Hongsheng, researcher Qian Chao of Zhejiang University, proposed for the first time a “green” knowledge inheritance neural network suitable for metasurface reverse design and assembly. The study was published in Light: Science & Applications under the title “A knowledge-inherited learning for intelligent metasurface design and assembly.”

The research team points out that unlike traditional “brick-to-brick” neural networks, where input-output parameters tend to be fixed and rigid, similar to building containerized houses with high flexibility and free assembly, the “board-to-board” algorithm proposed by the team gives the network recyclability and flexible composability (Figure 1a).

Figure 1: Schematic diagram of a knowledge inheritance neural network applied to metasurface reverse design. (a) Comparing house building to traditional neural networks and knowledge inheritance networks. In masonry construction, all bricks are stacked with mortar and fixed, while containerized buildings are built with removable “panel” components. (b) Knowledge inheritance paradigm. Similar to the “brick by brick” masonry building, traditional neural networks are indivisible, rigid, and single in function once established. In contrast, knowledge inheritance neural networks are suitable for multi-object-oriented and unfixed shape metasurface design tasks. It consists of two functional networks, INN and SNN. For a given “descendant” metasurface, the final network architecture can be synthesized by assembling the INN and dynamically adjusting the SNN.

Based on the special physical properties of metasurfaces, the proposed knowledge inheritance network is associated with the complex spatial information of metasurface structures, so as to realize the knowledge inheritance of the “parent” metasurface, and then construct a new “child” metasurface through free assembly. In other words, the method maps the assembly properties of metasurfaces in physical space to the assembly synthesis of neural networks. This paradigm breaks the stereotype that neural networks have long been suitable for predefined and single-shaped target design objects. Specifically, the knowledge inheritance network consists of two functional network modules, namely the inheritance neural network (full name INN) and the assembly neural network (full name SNN). In it, the INN is responsible for the reverse design of each “panel”/”parent” metasurface, and the SNN assigns subtasks to each INN as the deployer (Figure 1b).

The researchers benchmarked this paradigm using an assembly design of a large-scale non-periodic (Figure 2a) and three periodic “offspring” metasurfaces (Figure 4a), demonstrating its superior generalization ability. The advantages of this knowledge inheritance paradigm are obvious: mining, inheriting and recycling “parent” knowledge not only greatly reduces the design dimension, but also achieves the design task of multi-target scenarios.

Figure 2: Overall architecture and network structure of the knowledge inheritance paradigm. (a) Flowchart of the knowledge inheritance paradigm. To match the inheritance-assembly scheme, two networks (INN and SNN) are established, with INN responsible for the reverse design of each “panel” metasurface, which aims to explore the relationship between the global target electromagnetic response and the local electromagnetic response provided by each “panel” metasurface. In this paradigm, all local “panel” metasurfaces are built into a database containing seven local panels with different tilt angles (tilted in the direction), including 0°, ±10°, ±20°, -30°, and 45°. The irregular aperiodic metasurface consists of 49 local panels provided by panels A, B, F, and G. (b) Knowledge inheritance neural network structure. SNN is a dual-output network consisting of convolutional neural networks, and INN is established as a dual-input dual-output network with two modules. That is, a convolutional neural network module for reverse design and a physically auxiliary module for forward mapping (all numeric subscripts indicate the number of filters), two connected by an intermediate phase distribution.

Compared with transfer learning, knowledge inheritance networks have essential differences in basic mechanism and performance. Transfer learning is a mature algorithm derived from computer science when a pre-trained model is reused for another task. The basic operation process is to transfer a pre-trained neural network in the source task to assist in the training of the target task. However, this transfer method is extremely unstable and does not guarantee the success rate of transfer learning. Even the performance becomes worse compared to no transfer learning. In other words, transfer learning is like a “black box” that relies heavily on brute force attacks with features and lacks physical interpretability (Figure 3A).

Figure 3: Comparison of transfer learning and knowledge inheritance networks. (a) Transfer learning is an algorithm derived from computer science with unstable performance and feature-dependent brute force attacks that lack a reasonable explanation. (b) Knowledge inheritance learning is a unique method that often brings performance gains, with clear physical explanations and associations.

In contrast, knowledge inheritance learning is a unique and proprietary approach that can be viewed as a “white box” with physical connections between internally transferred knowledge (Figure 3b) and cannot be easily replicated in computer science. Due to the unique physical properties of metasurfaces, knowledge inheritance networks are associated with the complex spatial information of structures, which often brings performance gains, and realizes the echo of network synthesis in virtual space and metasurface assembly in physical space.

As the “loudspeaker” of satellites, the spaceborne antenna is an indispensable part of satellite communications. Traditional spaceborne antenna technology has considerable hardware cost, energy consumption, and computational complexity. Even with high accuracy, conventional electromagnetic solvers often rely on complex and lengthy numerical simulations for antenna inverse design.

In this study, the research team used the design of intelligent foldable metasurfaces to propose and experimentally demonstrate an innovative spaceborne antenna that is promising for future satellite communications. Flexible, ultra-thin and low-cost foldable metasurfaces can be mounted on satellite wings, removing the barriers of high hardware cost and computational complexity of traditional spaceborne antennas (Figure 4b). The experimental results clearly show the possibilities of satellite-satellite and satellite-earth communication, and the applied knowledge inheritance neural network also makes it possible to dynamically change the design goals during communication.

Figure 4: Periodic “offspring” metasurfaces and their assumptions for future satellite communications applications. (a) Assembly process of periodic foldable metasurfaces. Three “descendant” metasurfaces with different stretchable angles (±20°, ±10° and 0°) are assembled from panels D/E, B/C and A, abbreviated as subchild 1/2/3. Each “descendant” metasurface contains 16 panels (including 16×64 cells), 8 of which extend along the y-axis and 2 along the x-axis. (b) Schematic diagram of satellite communications based on intelligent foldable metasurfaces. As a flexible, ultra-thin and low-cost free beam steering device, the foldable metasurface can be mounted on the wing of a satellite to achieve free folding and stretching. Combined with the knowledge inheritance paradigm suitable for multi-target scenarios, foldable metasurfaces can realize satellite-to-satellite communication and satellite-earth communication. (Source: LightScience Applications WeChat public account)

Related paper information:https://doi.org/10.1038/s41377-023-01131-4

Special statement: This article is reproduced only for the need to disseminate information, and does not mean to represent the views of this website or confirm the authenticity of its content; If other media, websites or individuals reprint and use from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact the reprint fee, please contact us.



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button