Design and simulation methodology of surface acoustic wave diffractive neural network and its application in in-sensor computing

Abstract

With the development of deep learning, input data and the parameters in deep learning models surges demanding large computation resource. However, development of computation ability of a single chip slows down because the size of a single transistor almost reaches it physical limit. Aiming at processing abundant data from sensors, in-sensor computing, which combines neural network structures and sensor network, is advocated to reduce the data transformation between sensors and digital system.

Diffractive neural network (DFNN) is network structure utilizing the diffractive of wave. In the propagation of wave, Fresnel’s principle provides a foundational basis for understanding the diffractive processes, which exhibit mathematical parallels to the architectures of fully connected neural networks. This correlation posits that diffractive neural networks are capable of executing tasks traditionally associated with conventional neural networks, including classification and image recognition functions. The construction of DFNNs frequently leverages optical systems, capitalizing on the inherent benefits of light, such as its speed and parallelism. Additionally, alternative physical modalities, including radio frequency (RF) and ultrasound systems, have been explored for their potential application in DFNNs architectures.

Surface acoustic wave (SAW) devices, renowned for their high degree of integration, compatibility with biomedical applications, elevated frequency capabilities, and low noise profiles, are frequently employed as biosensors. The utility of SAW biosensors in the detection of cells and large molecular structures is primarily attributed to their ability to measure phase changes in the output signal, providing a sensitive means of analysis. One prevalent method of configuring SAW sensors for specific detection tasks involves the incorporation of antibody-antigen interactions. This approach leverages the specificity of the immune response, wherein antibodies bind to their corresponding antigens, to facilitate the detection of various biological entities. Such a mechanism underscores the adaptability of SAW devices in biosensing applications, offering a promising avenue for the development of sensitive and selective diagnostic tools.

The contributions of this thesis are in two aspects. Firstly, this work posits that Surface Acoustic Wave (SAW) systems can be innovatively applied in the construction of diffractive neural network. The inherent complexity and anisotropic properties of SAW present significant challenges in terms of calculation and simulation, distinguishing it from the more straightforward nature of optical systems. Through the application of finite element analysis, this research successfully navigates the complexities involved in the simulation and design of SAW-based systems, thereby facilitating their use in DFNN architectures.

Secondly, this thesis introduces a novel integration of SAW diffractive neural networks with SAW biosensors, aiming at the realization of sensor computing tasks. This proposed combination leverages the sensitive detection capabilities of SAW biosensors within the computational framework of DFNNs, specifically targeting the detection of conditions such as depression. This interdisciplinary approach not only underscores the versatility and potential of SAW technologies in both neural network construction and biosensing but also opens new avenues for the application of SAW-based systems in medical diagnostics and mental health monitoring.