Deep Submicron CMOS Reliability Concerns: TDDB and Tunneling Leakage

• Main Authors: Huan-Tsung Huang, 黃煥宗
• Other Authors: Prof. Ming-Jer Chen
• Format: Others
• Language: en_US
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/60078590233930560322

博士 === 國立交通大學 === 電子工程系 === 88 === This dissertation presents modeling analysis on two reliability concerns of deep submicron CMOS gate dielectrics : time-dependent-dielectric-breakdown (TDDB) and tunneling leakage. First, percolation methods were adopted in modeling TDDB. Second, tunneling leakage through oxide, both in the channel and the gate-drain overlapped regions were formulated and discussed. Since TDDB is a behavior combining random trap generation and path formation from one side to the other, the whole model should also contain two distinct parts in series relation, and with the randomly generated traps as the linkage. To uniquely interprete the experimental, statistical results, two most commonly cited percolation models, i.e. cell-based analytical model and sphere-based Monte Carlo model, were compared and were found to be correlative. This will lay an additional restriction on the choice of parameter values but sustain the consistency by adopting either one of the above models. \indent Although sphere-based Monte Carlo simulation successfully explained the thickness and area dependence of intrinsic breakdown statistics, the large conputation efforts kept it from practical application. Empirical reproduction of Monte Carlo simulations in closed-form was thus motivated. All features including ultimate thickness prediction were preserved in the proposed model with only the calculation time greatly down to ''click-and-response''''. To futher deal with the extrinsic case that is far more important in a real manufacturing process concerning the early failure, the concept of ''effective oxide thinning'''' was adopted in the Monte Carlo sphere model. Different from the original introduction of ''effective oxide thinning'''' concept that was a one-on-one relationship with the lower breakdown quantities, breakdown affected by the ''effective oxide thinning'''' was also competing in nature, just like its intrinsic counterpart. With ''effective oxide thinning'''' described by the geometric parameters as well as the percentage of occurrence among samples, the extended Monte Carlo sphere model now provides a direct way of modeling the competing nature in both the defective zone and the defect-free one. As a result, reproduction of the extrinsic or ''B'''' mode characteristics can create a new picture of defective severity as well as its uniformity across wafers and lots. In addition, sample-size-limited characteristics (detailed in chapter 4) as clarified by the improved model suggest that a care be taken when evaluating extrinsic TDDB data in a real manufacturing process. Again, for practical consideration, formulating both intrinsic and extrinsic statistics in closed-form was motivated and was achieved by applying the novel two region scheme (chapter 4) with each region treated intrinsically (chapter 3) and combining them through the competing risk formulism (detailed in chapter 5). In addition to providing underlying physical picture, the proposed model was computational efficient, free from the sample size limited problems encountered during the extrinsic TDDB evaluations, and, the most attractively, its easy adoption of local acceleration effects (field, temperature, etc.). Hole direct tunneling were found to dominate the valence electron tunneling in p+ poly gate pMOSFET''s with gate oxide thickness of around 3 nm or thinner. Modeling direct hole tunneling in an analytical way, we followed the steps that had been done on the electrons in the literatures. The proposed model could thus serve as a promising tools for sensitivly characterizing direct tunneling in oxides and can enable in-depth understandings of the subbands in the quantized inversion layer. On the other hand, as the gate oxide further down scaled to around 1.5 nm, the edge tunneling components of the gate tunneling leakage in off-state tended to overwhelm the channel ones. The model constructed with respect to the gate-drain overlapped region helped to extract the edge tunneling area of ~60 \AA times W (channel width). The experimental obervation also showed that the egde gate tunneling leakage was a single function of the gate-to-drain voltage difference (independent of the substrate bias), and the doping concentration adopted in the calculation was higher than 10^19 cm^-3. This high concentration implied the edge tunneling located in the deep extension region, consistent with the experimental observation.