*3.1. Design of Synaptic Device*

In designing the synaptic device, the focus was placed on successfully emulating biological neural system with Si compatibility, high scalability, high reliability, and high energy efficiency. In order to meet the requirements, various approaches were performed including embedding SiGe layer. There is a large difference in *E*<sup>g</sup> between Si and Ge and small difference in electron affinity (χ) so that most of the difference in energy bandgaps is transferred to VBO, which forms a hole QW in the SiGe region. Furthermore, SiGe is not only helpful in implementing potentiation mechanism but also in large current ratio between different weight states because its smaller *E*<sup>g</sup> has the effect of lowering the potentiation voltage compared with the all-Si case. Employing these features of SiGe, the SiGe layer can be used as short-term storage node, making the device more energy-efficient.

Figure 2a shows the block diagram schematically explaining the learning rule of human brain by Hebbian's law [29]. Hebbian's law effectively dictates the correlation-based plasticity in the biological nervous system where the connectivity between pre-neuron and post-neuron, i.e., the synaptic conductance is strengthened by repeated firing events of the pre-neuron. An increased number of pulses in a given time, or equivalently, an increased pulse frequency enhances the transition probability of the synaptic device from short-term to long-term memory. Figure 2b,c shows the energy-band diagrams in the channel direction and metal-oxide-semiconductor direction from gate 1 to gate 2, respectively. For the potentiation operation, BTBT is adopted as the primary mechanism considering device reliability, scalability, and energy efficiency (Figure 2b). As shown in Figure 2b, for a potentiation pulse, the valence-band electrons in the SiGe QW can see the empty states of the conduction band of the Si drain junction. As the result, holes are generated and effectively confined in the SiGe layer due to the large VBO between SiGe and Si. The locally confined holes by QW VBO give an effect of elevating the QW potential and increasing the channel conductance temporarily [30], which corresponds to the STP. Then, if the potentiation pulses are repeatedly applied to the transistor before the generated holes are annihilated by either recombination or diffusion, i.e., if the holes are accumulated and their amount exceeds a certain threshold value in the SiGe QW, LTP is introduced. The accumulated holes with the energies at the Fermi-Dirac distribution tail have higher probabilities of injection into the nitride charge-trap layer. Once the holes are trapped in the nitride layer, they do not vanish for long time, which establishes the LTP function. Moreover, work functions of those two gates are optimally adjusted to locate the BTBT site not in the vicinity of the right-side channel in order to prevent a soft potentiation and to store the generated holes at the right-side of the channel, which leads to a stable and reproducible LTP operation as shown in Figure 2c. By reflecting the aforementioned approaches, design of a synaptic device meeting the requirements is realized. mechanism considering device reliability, scalability, and energy efficiency (Figure 2b). As shown in Figure 2b, for a potentiation pulse, the valence-band electrons in the SiGe QW can see the empty states of the conduction band of the Si drain junction. As the result, holes are generated and effectively confined in the SiGe layer due to the large VBO between SiGe and Si. The locally confined holes by QW VBO give an effect of elevating the QW potential and increasing the channel conductance temporarily [30], which corresponds to the STP. Then, if the potentiation pulses are repeatedly applied to the transistor before the generated holes are annihilated by either recombination or diffusion, i.e., if the holes are accumulated and their amount exceeds a certain threshold value in the SiGe QW, LTP is introduced. The accumulated holes with the energies at the Fermi-Dirac distribution tail have higher probabilities of injection into the nitride charge-trap layer. Once the holes are trapped in the nitride layer, they do not vanish for long time, which establishes the LTP function. Moreover, work functions of those two gates are optimally adjusted to locate the BTBT site not in the vicinity of the right-side channel in order to prevent a soft potentiation and to store the generated holes at the right-side of the channel, which leads to a stable and reproducible LTP operation as shown in Figure 2c. By reflecting the aforementioned approaches, design of a synaptic device meeting the requirements is realized.

*Electronics* **2019**, *8*, x FOR PEER REVIEW 5 of 12

the energy-band diagrams in the channel direction and metal-oxide-semiconductor direction from gate 1 to gate 2, respectively. For the potentiation operation, BTBT is adopted as the primary

**Figure 2.** Operation principles of the synaptic device. (**a**) Hebbian's learning rule. (**b**) Energy-band diagram in the channel direction under the potentiation condition. (**c**) Energy-band diagram at the initial state and after potentiation state. The inset shows the band-to-band tunneling rate over a potentiation event. **Figure 2.** Operation principles of the synaptic device. (**a**) Hebbian's learning rule. (**b**) Energy-band diagram in the channel direction under the potentiation condition. (**c**) Energy-band diagram at the initial state and after potentiation state. The inset shows the band-to-band tunneling rate over a potentiation event.
