Stuck reset is an open‐circuit failure that occurs in phase‐change memory (PCM) after repeated reset and set operations, that is, endurance cycling. Stuck reset is majorly caused by phase‐change stress, which is the mechanical stress induced during a reset operation due to the density difference between the amorphous and crystalline phases of the phase‐change material. This indicates that a reduction in phase‐change stress may improve the endurance characteristics of PCM. Herein, a simulation technique for the calculation of phase‐change stress using a finite‐element software is proposed. Subsequently, a comparative study of the endurance of different PCM device architectures is performed. The results reveal that the self‐heating architecture exhibits superior endurance compared to the heater‐based architecture. Furthermore, the void locations in the experiments coincide with the most highly stressed locations in the simulation.

The variability of nonvolatile memory becomes more important as memory capacity increases. This is because, in general, individual device variability increases with scaling down. Using our simulation, we analyzed the intrinsic variability in phasechange memory switching originating from stochastic nucleation for self-heating and heater-based cell architectures. Differential equations of electrothermal and phase-field models were numerically solved in a fully coupled manner to simulate the amorphization and crystallization of the active material in the reset and set operations, respectively. Nuclei were seeded stochastically based on Poisson's probability. The phase distribution and cell resistance vary sample by sample owing to stochastic nucleation. The variability of reset resistance occurs because of nucleation during the falling time (FT) of the reset pulse, which is a short time of 10 ns. The resistance distribution was obtained by collecting data from 100 samples. All cells become reset by the 10 ns FT, with a narrow distribution of the reset resistance. As the FT increases, the resistance distribution becomes wider because some cells obtain more nuclei. As the FT becomes sufficiently long, the resistance distribution narrows again, and almost all cells are in the set state. The reset resistance is better fitted to the Weibull statistics than to the lognormal statistics. The heater-based cell showed less variability, whereas the self-heating cell consumed less energy during switching. Our simulation is useful in estimating the device performance and analyzing the variability as demonstrated in this report.