Reservoir computing (RC), a variant of recurrent neural networks (RNNs), is well-known for its diminished power consumption by unique deal with coaching the output weight and its superior efficiency in dealing with spatiotemporal data. Implementing these networks in {hardware} requires units with superior fading reminiscence habits. Not like filament-based two-terminal units, these counting on ferroelectric switching reveal improved voltage reliability, whereas three-terminal transistors present further lively management. HfO2-based ferroelectric supplies resembling Hf0.5Zr0.5O2 (HZO), have garnered consideration for his or her scalability and seamless integration with CMOS expertise. This research implements a RC {hardware} based mostly on MoS2–HZO built-in machine construction with enhanced spontaneous polarization discipline. By adjusting the oxygen emptiness focus, the units exhibit constant responses to each an identical and nonidentical voltages, making them appropriate for numerous RC purposes. The excessive accuracy of MNIST handwritten digits recognition highlights the wealthy reservoir states of the standard RC structure. Moreover, the influence of masks on RC implementation is assessed, showcasing the machine’s functionality for spatiotemporal sign evaluation. This growth paves the way in which for implementing energy-efficient and high-performance computing options.
