Ultrafast in‐memory computing and highly efficient deep neural networks driven by phase‐change memory materials with partially amorphous state transitions

Abstract Memristive materials (MMs) with adjustable conductance states are becoming the fundamental building blocks of artificial neural networks. However, challenges such as nonlinear weight conductance updates and the trade‐off between increasing melting uniformity and reducing solid‐to‐liquid transition time have impeded the demonstration of in situ learning on a large‐scale multiple‐layer memristive network. Here, we modulate the dynamic interactions between MMs and multiple‐pulse excitations in the melting‐based voltage‐pulse length diagram and temperature dependent mean‐square displacement. We utilize all four melting states (FMSs) to create an integrated framework for attaining rapid in‐memory computing and in situ deep neural network applications. We achieved near‐ideal R 2 value uniformity and a partial melting time below random‐access memory (RAM) device's switching times. Reversible switching using below‐RAM switching time pulses was also demonstrated. The FMS network discloses in situ learning capability and competitive classification accuracy on a conventional machine learning dataset. Simulations suggest that increasing the number of hidden neurons would further improve classification accuracy. Ab initio molecular‐dynamics simulations provide a clear insight into the melting kinetics in MMs and the structural origins of the prior complete melting‐facilitated decrease in output conductance. The memristive neural network represents a potential hardware platform for artificial intelligence, offering high speed and energy efficiency.