Enabling nondestructive readout in ferroelectric capacitors for computing-in-memory through interface engineering

Ferroelectric capacitors based on hafnium materials hold significant potential in nonvolatile memory and compute-in-memory computing, yet their on/off ratio and endurance under nondestructive readout remain critical challenges. This paper proposes an interface engineering strategy that effectively modulates the capacitive memory window at zero bias by inserting an ultrathin tungsten metal interlayer at the ferroelectric–dielectric interface. Experiments have demonstrated that devices with a 5 nm tungsten interlayer achieve a capacitive on/off ratio of 1.18 and exhibit exceptional endurance exceeding 108 cycles. This enhancement in performance is primarily attributable to the synergistic effects of charge modulation and localized field enhancement at the interface. A LeNet-5 neural network constructed using this device achieved a recognition accuracy of 95.16% in the MNIST handwritten digit recognition task. This research provides a viable device solution for developing high-performance, highly reliable edge computing hardware.