以二氧化鉿與氧化鋁作為閘極介電層之單層電荷儲存憶阻式鍺電晶體應用於類神經網路之研究

A Study of Single-Layer Charge Trap MemTransistor with the HfO2 and Al2O3-based Gate Stack on Germanium Substrate for Neural Network

易琴雅(Chin-Ya Yi)

指導教授 ： 簡昭欣

國立交通大學/電機學院/電子工程系所/碩士(2019年)

摘要

在此篇論文當中，我們製作出單層電荷儲存憶阻式鍺電晶體，在以二氧化鉿及氧化鋁為基礎的閘極堆疊(介面層/二氧化鉿或氧化鋁/氮化鈦)的結構。首先，我們先做出以二氧化鉿及氧化鋁當作電荷儲存層的兩種電容，可以看出兩者有明顯不同的電性表現，在二氧化鉿以及氧化鋁的條件中，分別有三種不同的厚度，可以看到越厚的電荷儲存層，能夠儲存的電荷越多，平帶電壓的平移量越大，然後我們比較了兩種材料的差異，電性顯示二氧化鉿的記憶保持力較差以及平帶電壓的平移量較小，此外，我們發現當我們給不同時間的脈衝，從100微秒開始以10倍增加，會得到均勻的平帶電壓位移量，而如果我們給相同時間的脈衝，平帶電壓的位移會達到飽和。
再來，我們製作出以二氧化鉿及氧化鋁當儲存層的單層電荷儲存憶阻式鍺電晶體，與二氧化鉿的例子相比，氧化鋁儲存層有更大的閾值電壓平移量，在相同的物理厚度下，它可承受的電壓比二氧化鉿電晶體可承受還來的大，這與我們之前電容的結果相符，且製作成電晶體後，氧化鋁電晶體的記憶保持力仍然優於二氧化鉿，在二氧化鉿儲存層中的電荷很容易且快速流失，因此我們推測氧化鋁有較好的抓取電子能力，更適合被採用於單層電荷儲存憶阻式電晶體。
最後，我們比較了兩種不同寫入/清除的量測模式，發現在採用相同時間脈衝來寫入的例子中，權重對脈衝數的圖線性度明顯比使用不同時間脈衝來寫入/清除的例子還要差，使用不同時間脈衝的這個方法較為適合應用於類神經網路加速，因此我們也是採用這種不同時間脈衝的寫入/清除方法來實現我們的應用。接著我們製作了2ⅹ2 的突觸陣列，以用來證明我們的電荷儲存式電晶體應用於類神經網路的可行性，我們藉由比較實際寫入電荷儲存憶阻式電晶體以及用matlab軟體所跑出的理想情形，可以發現實驗與理想狀況有很高的相似度，因此我們推論電荷儲存式電晶體應用於類神經網路是可行的，且綜觀以上的各種量測，發現氧化鋁電荷儲存憶阻式電晶體有較好的電荷保存力，較大的閾值電壓平移量以及較好的線性度，因此比單層二氧化鉿電晶體更適合應用於類神經網路加速。

關鍵字

二氧化鉿 ； 氧化鋁 ； 電荷儲存憶阻式 ； 類神經網路 ； 鍺電晶體

並列摘要

In this thesis, first, we fabricated Ge p-MOSCAPs based on the HfO2 and Al2O3-based gate stack before fabricating charge trap memtransistors. We compared Ge p-MOSCAPs based on three thicknesses of trapping layer. We found that the thicker trapping layer was; the larger VFB shift would be. Afterwards, we compared the difference of two materials. Electrical data shows the Al2O3 cases had larger VFB shift window and better retention. In addition, we found that there was uniform VFB shift after giving stepping pulse time increased by an order of 10 from 100 μs. However, the VFB shift would saturate when we gave identical time pulses.
Then, we fabricated charge trap memtransistors (CTMTs) based on the HfO2 and Al2O3-based gate stack. Compared to HfO2, Al2O3 trapping layer had larger Vth shift window. It could also afford higher voltage than HfO2 at the same physical thickness. The injection current of Al2O3 case was larger. Thus, there were more electrons injected and trapped in high-k layer, leading to Vth shift. The retention of Al2O3-based memtransistor was also better than that of HfO2-based memtransistor. Then, the read disturbance of Al2O3-based memtransistor was more stable than that of HfO2-based memtransistor. Therefore, we speculated that Al2O3 has better trapping ability, and it was more suitable as a trapping layer for single-layer charge trap memtransistor.
Finally, we compared two kinds of multi-programed/erased measurement with HfO2 and Al2O3-based charge trap memtransistors applied to the neural network. We knew the linearity between weight and pulse number in the case of identical pulses was worse than the case of stepping pulses. Although, when we programmed and erased with stepping pulses, the complexity of circuit design would increase. This method was still feasible for the neural network. Thus, we would adopt this PG/ER method to realize our applications instead of identical pulses. Afterward, we fabricated 2ⅹ2 synaptic array to demonstrate our charge trap memtransistors and introduced the measurement method. We also compared the result of our charge trap memtransistors and ideal case. According to the result, both of two cases still had high similarity to the ideal case. Therefore, charge trap memtransitor is a good implementation of synaptic device for neuro-inspired in-Memory computing due to its linearity on both weight-pulse number relation and input-output relation.. Nevertheless, Al2O3-based CTMTs had better retention, larger Vth shift window and better linearity. Al2O3-based CTMTs were more appropriate to be applied to the neural network as single-layer charge trap memtransistors.

並列關鍵字

Charge Trap MemTransistor ； HfO2 ； Al2O3 ； Neural Network ； Germanium Substrate

參考文獻

chapter1

[1] James M. Early, “Out to Murry Hill to Play: An Early History of transistors,” IEEE Transactions on Electron Device, Vol. 48, No. 11, Nov. 2011.

[2] G. E. Moore, “Crammimg more components onto integrated circuits,” Electronics, Vol. 38, pp. 114, 1965

[3] Kuzum, Duygu, et al. "Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing." Nano letters, vol.12, no. 5, pp.2179-2186, 2011.

[4] Strukov, Dmitri B., et al. "The missing memristor found." Nature, vol.453, pp.80-83, 2008.