Investigation of Phase-Change Synapse Architectures by Integrating Electro-Thermal and Phase-Field Models

Phase-change memory is a promising technology for synapse application in neuromorphic computing because it is the most matured among the nonvolatile memory technologies. The PCM utilizes Ge₂Sb₂Te₅(GST), a chalcogenide material, whose conductive crystalline and resistive amorphous phases correspond to data 0 and 1, respectively. The switching operation typically resembles heat treatment in material processing such as melt-and-quench for amorphization and annealing for crystallization via nucleation and growth. All the switching mechanisms are controlled by Joule-heating through electrical pulses. In this work, we constructed 3D PCM device simulation by integrating electrothermal and phase-field models in COMSOL Multiphysics. Two representative cell architectures are compared from the viewpoint of device performance and synapse application: one is self-heating wall (SHW), and the other is heater-based wall (HBW). Firstly, the device performance comparison including power consumption, on/off ratio, etc is conducted via recrystallization analysis by different falling-times in the range of 10ns and 100ns after reset operation. Secondly, the synapse characteristics are investigated, i.e., the conductance change data, which are obtained by applying a series of short and moderate current pulses for gradual crystallization after a reset pulse. From our simulations, we observe that the set operations by both melt-and-slow cooling and constant pulses (MSCP and CP) consume similar power. In a high-density memory array, the variability in devices is unavoidable. The MSCP with sufficient height can melt GST in all cells, which ensures successful operation. On the other hand, the CP with a fixed height may fail to crystallize in some cells due the variability. Therefore, the MSCP may be more beneficial than the CP. In case of synapse characteristics, the SHW is superior to the HBW because it consumes less power and shows better linearity and graduality in the synaptic weight.