Linear Conductance Modulation in Aluminum Doped Resistive Switching Memories for Neuromorphic Computing

With the advent of artificial intelligence (AI), automated machines could replace human labor in the near future. Nevertheless, AI implementation is currently confined to environments with huge power supplies and computing resources. Artificial neural networks are only implemented at the software level, which necessitates the continual retrieval of synaptic weights among devices. Physically constructing neural networks using emerging nonvolatile memories allows synaptic weights to be directly mapped, thereby enhancing the computational efficiency of AI. While resistive switching memory (RRAM) represents superior performances for in-memory computing, unresolved challenges persist regarding its nonideal properties. A significant challenge to the optimal performance of neural networks using RRAMs is the nonlinear conductance update. Ionic hopping of oxygen vacancy species should be thoroughly investigated and controlled for the successful implementation of RRAM-based AI acceleration. This study dopes tantalum oxide-based RRAM with aluminum, thus improving the nonlinear conductance modulation during the resistive switching process. As a result, the simulated classification accuracy of the trained network was significant improved.