| Summary: | 碩士 === 國立交通大學 === 電子研究所 === 108 === In this thesis, first, we fabricated Ge p-MOSCAPs based on the ZrO2-based gate stack, including: (1) single-layer p-ZrO2, (2) tA/tZ/tA, (3) pA/pZ/pA, (4) pA/tZ/pA, (5) pA/tZ/pA/tZ/pA, where t and p represented materials deposited by thermal ALD and PEALD, A was Al2O3 and Z was ZrO2. Electrical data showed that pA/tZ/pA gate stack case had the largest VFB shift window but poor retention. However, using pA/tZ/pA/tZ/pA gate stack could improve the retention but could degrade the VFB shift window slightly. Also, through XPS analyses, it was found that t-ZrO2/p-Al2O3 had larger conduction band offset than p-ZrO2/p-Al2O3 because p-ZrO2 contained the tetragonal and cubic crystalline phases and t-ZrO2 contained only the cubic crystalline phase and it had been proved by the first principle calculation that tetragonal ZrO2 had larger bandgap than cubic ZrO2.
Then, the pA/tZ/pA and pA/tZ/pA/tZ/pA gate stacks with PDA 400 oC were fabricated as the Ge charge trap memtransistors (CTMTs). Electrical data showed that the CTMT with the pA/tZ/pA gate stack had larger VTH shift window but the CTMT with the pA/tZ/pA/tZ/pA gate stack possessed better retention of charge. In addition, the larger the VTH was, the smaller the weight would be. It was found that the weight of CTMT with the pA/tZ/pA gate stack was easier to be depressed and potentiated than that with the pA/tZ/pA/tZ/pA gate stack.
In the end, through the adjustment of pulse time, the linearity between the weight and pulse number could be optimized. In addition, the weight of CTMTs in the both gate stacks would change about 0.2 after they were read for one hundred times. Next, we demonstrated the 2×2 synaptic array. It could be found that the weights of two kinds of CTMTs would affect the dependency between the post-neuron value and two pre-neuron values. To sum up, although the CTMT with the pA/tZ/pA/tZ/pA gate stack had better retention of charge, the weight of CTMT with the pA/tZ/pA gate stack could be depressed and potentiated more easily. Consequently, in our opinion, the CTMT with the pA/tZ/pA gate stack was more appropriate to be applied to the neural network.
|
|---|
