Programmable piezoelectric surface acoustic wave deep neural network with integrated memory function

Abstract

Over the past decade, artificial intelligence (AI) has experienced remarkable advancements, emerging as a fundamental driver of modern technological progress. The integration of AI into wireless radio frequency (RF) technologies, such as 6G networks, has significantly enhanced the intelligence and responsiveness of Internet of Things (IoT) edge devices. However, as the complexity of neural networks continues to escalate, the energy consumption and computational overhead associated with analog-to-digital conversion (ADC) and data transmission have far exceeded those required for computation, imposing a major constraint on AI applications in these edge computing scenarios. To address these challenges, novel electronic devices and computing architectures have been explored to sustain computational efficiency. Among them, analog wave-based computing, which integrates sensing, memory, and computation, presents a promising alternative. Unlike conventional digital architecture, analog wave computing exploits wave propagation to perform deep neural network (DNN) computations directly on analog signals, eliminating the need for ADC and mitigating the limitations of the "Von Neumann bottleneck" while enhancing computational parallelism. Direct electromagnetic (EM) wave-based DNNs have attracted particular interest due to their high processing speed. However, their large physical footprint and structural complexity pose significant challenges for integration into compact RF wireless communication system and remote sensing applications. To overcome these limitations, this thesis proposes a programmable DNN platform based on piezoelectric surface acoustic wave (SAW) technology. The reconfigurability of artificial neurons is achieved via the acoustoelectric effect, enabling dynamic SAW phase modulation through an applied bias voltage. A comprehensive investigation into SAW phase modulation mechanisms is conducted through finite element analysis (FEA) simulations and experimental validation. Three distinct phase-shifter architectures are examined: two reconfigurable designs leveraging carrier concentration modulation and velocity control, and one fixed-geometry variant utilizing metallic strip line variation. The interplay between semiconductor physics and acoustic wave propagation is quantitatively analyzed, providing critical design principles for high-precision phase control. Building upon this foundation, a hybrid SAW-memristor architecture is introduced by integrating an Ag⁺-based memristor with a carrier-migration-driven SAW phase shifter. This integration enables binary phase tuning with non-volatile bias retention, facilitating memory-enhanced neuromorphic computing. Leveraging the memory-enabled phase shifter, a binarized SAW-based neural network was both simulated and experimentally implemented for vector classification tasks. Both simulation and experimental results demonstrated successful vector classification, with the experimental implementation achieving an classification accuracy of 91.7%, highlighting the strong classification performance of the SAW DNN. These findings suggest that the proposed highly integrated, chip-scale, memorized programmable SAW DNN platform holds significant potential for applications in remote sensing, wireless signal identification, and other edge computing domains.