Linearly Programmable Two-Dimensional Halide Perovskite Memristor Arrays for Neuromorphic Computing

Neuromorphic hardware, which provides high-performance AI processing capability with low power consumption, is an attractive and challenging field designed to overcome the existing von Neumann computing systems. To implement high-performance training in neuromorphic hardware, it is essential to develop artificial synapses that exhibit linear and symmetric programmability with a bipolar operation, analog multi-states with a high dynamic range, a high yield, a long retention, a low variation, and a small footprint. To achieve these requirements, memristors, non-volatile memory devices that store data by their conductance, have been widely studied as artificial synapses. However, traditional memristors lack a reliable microscopic structure to confine ion migration during switching, resulting in commonly observed large variability (from device to device and switching cycle to cycle) and abrupt switching (instead of linear and symmetric programming). To address these issues, numerous approaches have been explored, such as modulating conductance by adding gate-terminals or optimizing programming schemes. Only limited success has been achieved so far, which, on the other hand, typically incurs substantial area, circuitry, time, and/or energy overheads.<br/>Recently, two-dimensional (2D) halide perovskites have arisen as a top candidate for artificial synapses due to their phase versatility, superior memristive properties, microstructural anisotropy in electrical and optoelectronic properties, and even excellent moisture resistance. Unfortunately, a common challenge in all memristors has also been identified in such halide perovskites, namely, asymmetric and nonlinear conductance change, which is a well-known roadblock for efficient training and accurate inference when such materials are used in neural networks.<br/>Here, we achieve highly linear and symmetrical conductance changes (α_p: 0.002, α_d: -0.0015) in Dion-Jacobson 2D perovskites, which were unachievable previously in 2D perovskites. We further build a crossbar array based on analog perovskite synapses, achieving a high (~100%) device yield, low variation (~1.85%) with synaptic weight storing capability, multilevel analog states with long retention (~10⁴ s), and moisture stability over 7 months. We explore the potential of such devices in large-scale image inference via simulations and show an accuracy within 0.08% of the theoretical limit. The remarkable device performances are attributed to the homogenous migration of halide vacancies by eliminating gaps between inorganic layers, confirmed by first-principles calculations and experiments. Due to the Dion Jacobson phase formed by changing large organic cations from monovalent to divalent ammonium cations (A′′A_n-1Pb_nX_3n+1, A′′ is divalent ammonium cation), two hydrogen bonds are formed between organic and inorganic layers, eliminating van der Waals gaps, resulting in homogeneous interfacial ion migration through the entire region of vertically aligned layers. Our materials design rule is generally applicable to other memristive material systems for achieving high-performance neuromorphic computing.