Realizing the Gestalt Principle of Closure with Ultra-High Denisty Three-Dimensionally Integrated Perovskite Nanowire Array Based Artificial Neural Network

The application of halide perovskite (HP) thin-film based resistive random access memories (RRAMs), in the domain of neuromorphic computing, has been hindered by the material and electrical instability of HP thin-films, leading to less than ideal performance metrics such as poor retention, jittery conductance levels and sub-par endurance. To overcome these limitations, we engineered a unique switching matrix that substitutes the thin-film design with vertically arranged, three-dimensional, high-density monocrystalline HP nanowires (NWs), which are embedded in a porous alumina membrane (PAM) and enclosed between electrical contacts.<br/>Specifically, we developed robust, lead-free, HP based artificial synapses utilizing monocrystalline, ultra-high density (10¹⁰ cm^-2) methyl ammonium bismuth iodide (MA₃Bi₂I₉ or MBI) nanowires (NWs) rooted in a nano-engineered PAM forming a stable structure that enhances the synaptic and neuromorphic device performance. This innovative approach offers outstanding passivation, providing the necessary electrical, thermal and material stability to the environmentally sensitive HP by significantly minimizing surface diffusion pathways, thus preventing moisture-induced damage.<br/>The MBI NW devices exhibited two key behaviors: short-term plasticity (STP) and long-term potentiation (LTP). These behaviors were elicited in response to input sensory electrical pulses of varying amplitude, duration, and interval. The devices transitioned from STP to LTP mode with an increased number of stimulating pulses and in the LTP mode showed a response to the rate and number of incoming LTP pulses and the pre-/postsynaptic delay between pulses.<br/>The origin of the conductivity change in the monocrystalline MBI NWs was traced through first principle ab initio molecular dynamics simulations. The simulations revealed that the rotation of the MA+ clusters in the MBI crystal structure results in charge transfer between MA⁺ and Bi₂I₉³⁻, leading to the observed plasticity or gradual conductivity change.<br/>The MBI NW devices demonstrated at least 40 accessible conductance states which could be retained for more than 10⁵ seconds. These states showed minimal jitter, with a variation (2σ where ‘σ’ is the standard deviation/mean) of less than 10%, and an endurance of approximately 10⁵ cycles.<br/>We further developed a single-layered artificial neural network (ANN), utilizing MBI NW synapses as the unit blocks. This ANN was trained using synaptic weights obtained from delta rule-based simulations for a letter recognition scheme. The ANN successfully discriminated between fragmented and non-fragmented letters, demonstrating the potential for pattern recognition tasks emulating human cognition. Additionally, through the use of the MBI NW-based ANN, it was experimentally shown that as more parts of a single letter are missing, the ability of the MBI NW artificial synapses to differentiate and maintain synaptic connectivity decreases, similar to our human perceptual tendencies.<br/>In conclusion, this study not only pushes the boundaries of halide perovskites in neuromorphic processing tasks but also exemplifies the Gestalt principle of closure. This principle posits that humans tend to perceive incomplete or fragmented forms as complete, thereby bridging the gaps in sensory input. The successful implementation of this principle in a letter recognition scheme using an MBI NWs-based ANN demonstrates the potential for neuromorphic hardware to perform complex processing tasks that emulate human cognitive processes.