Ferroelectric-Assisted van der Waals Charge-Dynamics Reservoir Arrays for Spatiotemporal Computing.

Abstract

Physical reservoir computing offers an energy-efficient framework for spatiotemporal processing by exploiting nonlinear dynamics and fading memory. Two-dimensional semiconductors provide intrinsic interfacial charge-trapping dynamics as a potential source for such behaviors. However, purely defect-driven reservoirs suffer from limited tunability and variability. Here, we demonstrate a ferroelectric-assisted charge-trap memory (FA-CTM) that transforms inherently random interfacial charge dynamics into a highly controllable, designable physical reservoir. By introducing a ferroelectric underlayer and engineering voltage partitioning via capacitance matching, the FA-CTM is intentionally driven in a volatile regime that strengthens nonlinearity while preserving reproducible fading-memory behavior. This architecture achieves a 51% enhancement in the memory window together with a 12-fold increase in on-current compared to pristine dielectric counterparts. Through an industry-academia partnership, we validated array-level scalability using 8-inch wafer-scale MOCVD-synthesized MoS₂ channels integrated via dry transfer processes. The result exhibits exceptional uniformity, with average reservoir state output-current variation minimized to 3.11%, compared to 15.57% in non-ferroelectric controls. Consequently, the FA-CTM reservoir achieves 90.18% accuracy on EMNIST classification and supports a multitasking pipeline for video-based object recognition and trajectory prediction. These results establish ferroelectric-assisted field control as a practical method to co-optimize tunability and uniformity in charge-trap memories, providing a robust foundation for spatiotemporal processing.