We report a monolithic three-dimensional integration of dendritic neural network (M3D-DNN) with memristors-based artificial synapse, dendrite and soma on top of Si-based CMOS logic. The Si CMOS layer served as control logic fabricated in foundry. A 1k-bit artificial synaptic array was built with HfO $_{2} -$based nonvolatile memristors to implement computing-in-memory (CIM). In addition, TiO $_{x} -$based memristive artificial dendrite and NbO $_{x} \mathrm{N}_{y} -$based memristive artificial soma were adopted to implement the dendritic neuron (DN) layer to process postsynaptic signals. Both the CIM and DN layers were fabricated using a BEOL-compatible process. The structural integrity and proper function of each layer in the M3D-DNN were verified. Our work demonstrates a promising architecture to efficiently implement bio-plausible artificial neural networks (ANNs).

With the rich internal ion dynamics, memristor-based neuromorphic computing emerges as a non-von Neumann computing paradigm to mimic biological neural networks and achieve high energy efficiency. However, to implement large-scale memristive neural networks, the reliability issue of memristive devices, including artificial synapse, dendrite, and soma, should be properly addressed. In this paper, recent works investigating the physical mechanisms and optimizations of memristive device reliability are presented. In particular, the relaxation effect of $\boldsymbol{\text{HfO}_{\mathrm{x}}}$ -based artificial synapse is alleviated by using a ternary oxide as the thermal enhance layer, the device yield of $\boldsymbol{\text{TiO}_{\mathrm{x}^{-}}}$ based artificial dendrite is improved by proper material selection and interface engineering, and the device variability of $\boldsymbol{\text{NbO}_{\mathrm{x}}}$ -based artificial soma is reduced by nitrogen doping. Furthermore, a bio-inspired dendritic neural network with these three fundamental memristive devices is constructed and simulated to analyze the influence of device reliability. Using these optimized devices, the classification accuracy of the street-view house number dataset can be improved by up to $\sim$ 60%. The quantitative requirements of device reliability metrics are also provided as a guideline for future neuromorphic system design and implementation.

Mott memristors have been considered as a promising candidate to implement artificial neurons for neuromorphic computing thanks to their low-power consumption and superior scalability. However, the large variability and poor reliability hinder their large-scale applications. The complex working mechanism associated with the thermoelectric coupling in the correlated oxides such as niobium oxide (NbOx) has led to the lack of a physics-based model to guide device optimizations. In this work, we present a microscopic model of NbOx-based Mott memristor and investigate the evolution of atomic configuration via a real-time scale kinetic Monte-Carlo simulation involving multiple physical processes. We elucidate the relationship between the I-V characteristics and the oxygen stoichiometry. We further reveal that the low-yield issue originates from the oxidation of NbO2 phase in air and the poor reliability correlates with the migration of oxygen vacancies. We hence propose to improve the device performance by introducing a Si3N4 passivation layer and N doping. The optimized devices exhibit excellent endurance of more than 108 cycles with significantly reduced variability and low operation voltage. Both oscillation neuron and leaky integrate and fire (LIF) neuron are experimentally implemented using the optimized Mott device, which could serve as a highly reliable artificial neuron with low variability and excellent endurance for large-scale neuromorphic computing systems.