A Reconfigurable Ferroelectric Acoustic Computational Primitive for Analog In-Memory Computing

The growing computational demands of artificial intelligence (AI) have renewed interest in emerging nonvolatile memory (eNVM) crossbar array technologies for in-memory computing to reduce the energy cost of moving data between processors and memory. Here we report a computational primitive that merges ferroelectric memory and acoustic resonators in a single platform: Ferroelectric memory Embedded in Nanoscale Acoustic resonators (FENA). We experimentally demonstrate over 100 nonvolatile analog admittance states arising from partial ferroelectric polarization switching in lithium niobate resonators. The devices exhibit highly symmetric potentiation and depression characteristics, which directly impact computation accuracy, favorable dynamic range and linearity, and low projected read and write energies after scaling. We further propose a radio frequency (RF) peripheral readout architecture that extracts frequency- and phase-sensitive effective conductance states to offer an expanded state space and hardware-level reconfigurability. System-level evaluation using the NeuroSim multi-layer perceptron (MLP) simulator demonstrates the suitability of FENA for both online and offline neural-network training, with scaled devices achieving competitive 84% and 92% classification accuracies, respectively.