Metasurface Achieves Parallel Classification Mimicking Neural Network Computation.

A metasurface realising a physical neural network concurrently performs two independent classification tasks via parallel processing. This is achieved by multiplexing carrier frequency through temporal modulation of embedded waveguides, overcoming limitations of single-task physical wave-based systems and enabling complex computation.

The development of physical systems capable of complex information processing remains a significant challenge. Researchers are increasingly exploring the potential of metamaterials – artificial structures engineered to exhibit properties not found in nature – to mimic the function of neural networks. A team led by M. Mousa, M. Moghaddaszadeh, and M. Nouh, all from the Department of Mechanical and Aerospace Engineering at the University at Buffalo (SUNY), detail a novel approach to this problem in their paper, “Analog dual classifier via a time-modulated neuromorphic metasurface”. Their work addresses the limitation of single-task operation in existing physical neural networks by demonstrating a metasurface capable of performing two independent classification tasks concurrently, achieved through the precise manipulation of wave propagation via temporal modulation.

Mechanical Metamaterials Enable Concurrent Analog Computation

A mechanical metamaterial has been developed capable of performing parallel computation through the manipulation of elastic waves, establishing a foundation for future energy-efficient computing architectures. Researchers designed a dual-classifier metasurface, validating both theoretical predictions and experimental results to demonstrate its effectiveness.

Neural networks, computational systems inspired by the biological brain, comprise interconnected nodes (‘neurons’) organised in layers. These networks ‘learn’ by adjusting the strength of connections between nodes. In this mechanical system, the equivalent of these trainable parameters are tunable phases within the metasurface, achieved through precise control of resonant structures. These phases modulate the behaviour of elastic waves – vibrations that propagate through solid materials – as they traverse the device.

Previous physical implementations of wave-based computation typically address only a single task, a limitation stemming from the sequential nature of wave propagation. To circumvent this, a frequency multiplexing strategy was implemented, encoding different classification tasks onto different frequencies of elastic waves. By simultaneously launching waves at these distinct frequencies, the metasurface performs both classifications in parallel.

The system’s architecture incorporates embedded waveguides – structures that guide the propagation of waves – with precisely prescribed temporal modulations critical for enabling frequency multiplexing and ensuring accurate parallel processing. Researchers addressed constraints imposed by wave propagation in finite media, paving the way for more complex and versatile physical computing platforms.

Unlike traditional electronic computers, these wave-based systems offer potential advantages in energy efficiency and speed for specific computational tasks. Further development could lead to novel architectures for signal processing, pattern recognition, and other applications requiring parallel computation. This research represents a convergence of materials science, physics, and computer science, establishing a new wave-based computing paradigm.

The ability to perform parallel computation within a physical system opens up possibilities for energy-efficient and high-speed processing, potentially impacting fields such as sensor networks, robotics, and real-time data analysis. The demonstrated principle of frequency multiplexing offers a promising route towards building more complex and versatile physical neural networks capable of tackling increasingly challenging computational tasks.