以電阻式記憶體為基礎之?電子突觸元件與精簡模型

Abstract

大腦被視為世界上功能性最佳的「產品」。因此，研究人員致力於開發能模仿大腦低功耗、高密度、平行處理特性的仿神經型態系統。在仿神經型態運算中，突觸主要的功能在於學習以及記憶，扮演關鍵性的角色。因此，欲達到仿神經型態系統的首要目標即是模仿生物突觸的行為，製作出人工電子突觸元件。目前已經有許多科學家做出人工電子突觸元件，其中，電阻式記憶體，有時候亦被稱為憶阻器，被認為是最具潛力的候選人之一，因為它單位面積小、操作速度快、低功耗的優點能滿足電子突觸應用的要求。但是，燈絲型態的電阻式記憶體有關鍵性的缺陷待克服。其一，它的寫入以及抹除有很大的變異性，不利於電路設計。其二，當製作成三維陣列的產品時，易產生漏電流路徑導致讀取錯誤。 在此篇論文中，我們使用非燈絲型的電阻式記憶體－Ta/TaOx/TiO2/Ti為人工電子突觸，可應用於未來仿神經型態運算。此非燈絲型的電阻式記憶體擁有很多優點，包含無須額外的生成操作、自我限流、自我整流、和主動元件的製程匹配、和漸進式寫入及抹除。而且，因為操作機制不同，非燈絲型電阻式記憶體具有極小的寫入及抹除變異性。此外，在此論文中實現此非燈絲型電阻式記憶體於三維結構中，因為自我整流特性可有效抑制漏電流路徑，避免讀取錯誤。 本研究於第二章中利用二維的人工電子突觸元件展示基本的突觸特性。其中，示範的動作電位和時序相依可塑性(STDP)是利用仿生物的動作電位，此模擬將比一般研究所用的方型脈衝波更為仿生。而其增益性與抑制性學習曲線呈現穩定性、類比性且可一再重複的學習特性。至於維持記憶的方面，依據不同的條件，可持續幾分鐘至幾小時不等，可視為短期記憶至長期記憶的轉變過程。上述結果驗證了Ta/TaOx/TiO2/Ti人工電子突觸元件具有優異的生物突觸特性，適合利用於建構未來仿神經型態系統。 於第三章中，我們根據Ta/TaOx/TiO2/Ti元件本身的操作機制和憶阻器的廣泛定義，建構了一個以物理為基礎的解析解精簡模型。精簡模型可利用於電路特性模擬，為日後設計大型仿生運算系統所不可或缺的。我們提出的模型不但可以很好地模擬小操作電壓範圍之下的增益性及抑制性，且模擬的動作電位和時序相依可塑性也和量測結果吻合，確認了此精簡模型的可信度。 於第四章中，我們進一步將人工電子突觸元件實現於高密度三維結構中，三維人工電子突觸陣列為日後建構如人腦般高密度三維神經網絡之關鍵。我們所實現的三維人工電子突觸元件不但具有如二維元件般優異的生物突觸特性，包含在持續幾百個波形操作增益及抑制特性過程中維持穩定之變化範圍，以及可藉由輸入波型數目調變增益性及抑制性範圍的特性。且因為非燈絲型Ta/TaOx/TiO2/Ti元件具有操作電流可微縮之特性，在奈米尺度元件中可提供極低的操作電流，因此單一動作電位只需要fJ的能量損耗，與生物突觸的功耗表現接近，對比過去人工電子突觸元件研究中，最低的能量耗損僅到達幾百fJ等級，以上種種優勢讓本論文中提出的元件近乎和生物突觸一樣功能強大。

Human brain has been considered to be the most powerful “device” in the world. To realize low-power, high-density, and efficient parallel computation, researchers are actively investigating neuromorphic systems that operate in a similar manner of our brains. In the neuromorphic computation, synapse plays an important role in learning and memory. Therefore, to emulate the neuromorphic system, constructing an artificial electronic synapse is a critical step. Among all kinds of synaptic devices, resistive random access memory (RRAM), sometimes also called memristor, is one of the most promising candidates because of its 4F2 cell size, fast switching speed, and low power consumption. However, the variations of the SET and RESET operations in filamentary RRAM have been a critical issue to overcome. More importantly, it suffers from the sneak current issue when implementing a high-density three-dimensional (3D) array network,. In this thesis, we investigate a non-filamentary Ta/TaOx/TiO2/Ti RRAM as an artificial electronic synapse for future neuromorphic computation. The Ta/TaOx/TiO2/Ti double-layer RRAM has many outstanding advantages, including forming free, self-compliance, self-rectification, usage of fab-friendly materials, and gradual SET and RESET characteristics. Because the switching mechanism is different from that in filamentary RRAM, the variations of SET and RESET operations are greatly mitigated. Moreover, the proposed device can be used in 3D vertical arrays because the self-rectifying characteristics suppress the undesired sneak current. In Chapter 2, we summarized and demonstrated the basic synaptic characteristics in our two-dimensional (2D) synaptic RRAM. To mimic biological synapses. The spike-time-dependent-plasticity was measured using action-potential-like spike rather than regular square pulses. The potentiation and depression training processes demonstrates monotonic and analog synaptic weight changes and reproducible learning behaviors. Memory retention in our device lasts for minutes to hours, representing the transition from short-term memory to long-term memory. In Chapter 3, we construct an analytical compact model of our Ta/TaOx/TiO2/Ti device based on the general definition of memristors and the physical switching mechanism of the device. The compact model can be used in circuit-level simulation, and is essential for building large-scale neuromorphic computing systems. The proposed model was verified by the good agreement with the experiments in a finite programming voltage range of potentiation and depression and STDP. In Chapter 4, we further construct a 3D vertical synaptic device that is critical in constructing future high-density 3D neuron networks similar to our brains. In addition to excellent synaptic characteristics similar to its 2D counterpart, such as stable and tunable potentiation and depression training, the energy consumption per spike can be as low as a few fJ because of the extremely low operating current in the nanoscale device. The energy consumption is comparable to biological synapses and much lower than previous artificial synapses with an energy consumption of at least a few hundred fJ. With all of the advantages reported in this thesis, the proposed device could be as powerful as biological synapses and shows promising potential for future neuromorphic systems.