A New Read Scheme for Alleviating Cell-to-Cell Interference in Scaled-Down 3D NAND Flash Memory

: In this paper, we investigated the cell-to-cell interference in scaled-down 3D NAND ﬂash memory by using a Technology Computer-Aided Design (TCAD) simulation. The fundamental cause of cell-to-cell interference is that the electric ﬁeld crowding point is changed by the programmed adjacent cell so that the electric ﬁeld is not su ﬃ ciently directed to the channel surface. Therefore, the channel concentration of the selected cell is changed, leading to a V th shift. Furthermore, this phenomenon occurs more severely when the selected cell is in an erased state rather than in a programmed state. In addition, it was conﬁrmed that the cell-to-cell interference by the programmed WL n + 1 is more severe than that of WL n − 1 due to the degradation of the e ﬀ ective mobility e ﬀ ect. To solve this fundamental problem, a new read scheme is proposed. Through TCAD simulation, the cell-to-cell interference was alleviated with a bias having a ∆ V of 1.5 V from V read through an optimization process to have appropriate bias conditions in three ways that are suitable for each pattern. As a result, this scheme narrowed the V th shift of 67.5% for erased cells and narrowed the V th shift of 70% for programmed cells. The proposed scheme is one way to solve the cell-to-cell interference that may occur as the cell-to-cell distance decreases for a high stacked 3D NAND structure.


Introduction
Since the release of the 3D NAND flash product for the first time in 2014 [1], the bit density could be increased rapidly by increasing the number of WL (Word line) stacking layers, which starts at 24, came up to 128 [1][2][3][4][5][6], and it is unpredictable how many it will increase to fulfill the explosive growth of data demand. In this situation, while increasing the number of stacked WL layers, the mold stacking height should be scaled down due to the limitations of the channel hole etching process. However, if the gate and space length are reduced through the height being scaled down, the memory characteristics change, and it leads to reliability issues [7][8][9][10][11][12][13][14]. One of the issues is due to the junction-less poly silicon channel of 3D NAND structure. In effect, the channel of the main cells is based on the ultra-thin body (UTB) structure, which is not connected to the substrate [15,16]. Therefore, during the read operation, the selected cell is affected by the V read (read voltage) of the adjacent cell, which leads to cell-to-cell interference [11,12,17]. Since the cell-to-cell interference is different from the conventional FG (floating gate) of a planar 2D NAND structure [18][19][20], research on the cell-to-cell interference in 3D NAND has not been investigated yet. Therefore, in this paper, when scaled down in a 3D NAND structure, cell-to-cell interference according to each pattern was studied using TCAD simulation (Synopsys Sentaurus TM ). This is not only to investigate the interference according to the bias condition of the adjacent cells but also to consider the program state of all cells in the NAND string. After that, we proposed a new read scheme using a local bias condition that can alleviate the cell-to-cell interference.

Simulation Set-Up
In this paper, TCAD simulation was performed to analyze the cell-to-cell interference of 3D NAND flash memory. The 3D structure used is shown in Figure 1a. The NAND string is a 16-layer structure, in which a BL (Bit-Line), SSL (String Select Line), GSL (Ground Select Line), and CSL (Common Source Line) are constructed. Typically, both the gate and space lengths are scaled down at the same time in 3D NAND flash memory. However, in this simulation, the gate length was set equal to 30 nm for all structures, and the space length was reduced by 5 nm from 30 nm and scaled down to 10 nm. The reason for this setting is if the gate length is changed, a difference in both the gate controllability and the charge trap density occurs. We set WL9 as a selected cell of the read operation and confirmed cell-to-cell interference using the states of adjacent cells WL n−1 and WL n+1. Figure 1b is the equivalent circuit and bias conditions to read through a 3D NAND string. Figure 1c is the pulse diagram of the program and read operation. The program voltage is 20 V and the pass voltage is 6 V with 3.3 V applied to the SSL in the selected bit line. The read voltage is 6 V with 3.3 V applied to the SSL/GSL. The selected cell bias is swept from −5 V to 5 V to verify the cell state and the BL voltage is set to 1 V. The list in Table 1 shows patterns for investigating interference in selected cells (WL n ) according to the state of the adjacent cell. We have listed all patterns for three cells. Patterns A, B, C, and D are used to identify interference when the selected cell is in the erased state, and the patterns E, F, G, and H are configured to identify interference when the selected cell is in the programmed state. Lastly, in the TCAD simulation, the physics model in poly-silicon used a doping dependence model and high-field saturation model for mobility. Furthermore, the recombination model used the doping dependence of the SRH (Shockley-Read-Hall) lifetime model, respectively. In addition, the poly-silicon channel is affected by the GBT (Grain Boundary Trap) for memory operation [21]. In this simulation, we excluded this effect to avoid complexity and only considering the cell-to-cell distance effect.   Figure 2 shows the V th shift caused by cell-to-cell interference with scaled down in 3D NAND flash memory according to each pattern. As the space length decreases, it can be seen that the interference increases regardless of the pattern type. However, the interference is more severe than the programmed state when the selected cell is in the erased state. In addition, patterns D and H, where WL n+1 and WL n−1 are programmed, have the most serious cell-to-cell interference; however, we should note that the interference caused by programmed WL n+1 is more severe than that of WL n−1. First, the cell-to-cell interference caused by the reduction in the space length was examined through the process of changing the state from pattern A to B, and E and F. Figure 3a,b show the conduction band at the channel surface when the selected cell is turned on (V verify = V th of selected cell) in the read operation. In Figure 3a, which has both gate and space length and a 30 nm structure, it can be confirmed that the conduction band at pattern B is relatively higher than that of pattern A at the WL n−1 channel region despite the V read bias. This acts as a source resistance during the read operation, but the interference is considered negligible because there is no change in the conduction band in the selected cell channel region nearby. In the case of patterns E and F, there is no change in the conduction band.

Cell-To-Cell Interference
However, the space length is scaled down to 10 nm in the 3D NAND structure in Figure 3b; when the read operation is changed from pattern A to B, the raised conduction band of the programmed WL n−1 overlaps the selected cell channel area. As a result, the conduction band distortion occurs. Finally, when comparing the conduction bands of patterns E and F in a 10 nm length structure, it was found that the distortion is relatively negligible compared with patterns A and B. To examine these causes in detail, Figure 4 confirmed the electron density at the channel area during the read operation at each pattern. Figure 4a,b shows the electron density contour according to the pattern in the space length, which is 30 nm when the selected cell is turned on during the read operation. As WL n−1 was programmed as the pattern changed from A to B, the electron density of the inversion layer by V read was lower than that of pattern A. However, it does not affect the channel area of the selected cell. This is because it maintains a strong inversion state nearby the selected cell. As a result, the verify voltage in Figure 4b is −0.761, which is as high as 0.067 V than Figure 4a. Likewise, in Figure 4c,d, when the space length is 10 nm, the electron concentration contour changing from pattern A to B can be confirmed. Compared with Figure 4a,b, and since the space length is shortened to 10 nm, Figure 4d shows that the insufficient electron concentration in the WL n−1 channel region overlaps the channel concentration of the selected cell. As a result, it maintains the light inversion state at a nearby selected cell. Therefore, the verify voltage in Figure 4d is −1.860, which is as high as 0.279 V in Figure 4c. Lastly, in Figure 5e,f, when the space length is 10 nm, the electron concentration contour changing from pattern E to F can be confirmed. As in Figure 4b, it can be confirmed that the electron concentration of the selected cell is not affected due to the strong inversion. As a result, the verify voltage in Figure 5f is 2.540, which is as high as 0.04 V compared with Figure 5e. Therefore, if the space length is short enough, cell-to-cell interference is negligible in the programmed selected cell.  Specifically, to investigate the difference of electron concentration at each state, Figure 5a shows the electric field contour and direction generated during the read operation. The electric field crowding point occurs near the nitride trap layer of the WL n−1 due to the trapped electron charge and near the selected cell gate region due to the low verify voltage compared to the read voltage. As a result, the electric field generated through the V read voltage applied to WL n−1 is directed to the crowding point, but the fringe field can bypass the channel region through sufficient space length; as a result, it can be confirmed that the electric field is directed to the channel surface leading to strong inversion nearby the selected cell, as shown in Figure 4b However, as shown in Figure 5b, when the space length is short, there is no distance that the electric field can bypass. So, it can be confirmed that the electric field is not directed to the channel region leading to light inversion as shown in Figure 4d. Finally, Figure 5c shows the electric field in Figure 4f. Compared to Figure 5b, the verify voltage increased due to the programmed state; therefore, the electric field crowding point was changed from the gate to the nitride trap layer. As a result, the fringe field can reach the channel surface region, leading to strong inversion.
In conclusion, due to the difference in field direction, a difference in electron concentration in the channel surface region occurs. Through the cell-to-cell interference due to the change from pattern A to B and E to F, we confirmed that the interference with WL n−1 is the inability to form a junction due to an insufficient electric field.
However, as shown in Figure 3, the cell-to-cell interference due to the change from pattern A to C and E to G is more severe. Figure 6d shows all the conduction bands at the channel surface in patterns A, B, C, and D at a space length of 10 nm in the 3D NAND Structure during the read operation. The conduction band is significantly increased by the programmed WL n+1 compared with the programmed WL n−1 . The conduction band slope between WL n+1 and WL n becomes smooth, leading to conduction band distortion. Through this point, cell-to-cell interference from pattern C is more severe than pattern B. and this conduction band distortion effect inferred a significant drop in the electric field.
In effect, during the read operation, drift current occurs through the V bl and V read , but the electric field due to V read is weakened by the programmed WL n+1 . Therefore, as shown in Figure 6a, while in the gate and space length of a 30 nm structure, there is no change in effective mobility in patterns C and D because sufficient fringe fields by V read are directed to the space region, when the space length is 10 nm as shown in Figure 6b, effective mobility is significantly degraded because the fringe field could not be directed to the channel surface. Therefore, as shown in Figure 6c, due to this mobility degradation, electrons flowing through the field over the channel increase SRH recombination, leading to lowering the channel concentration and more severe cell-to-cell interference than programmed WL n−1 induced interference.
Lastly, in simple terms, pattern D has the greatest interference because it affects both WL n−1 and WL n+1 interference. As for the patterns E, F, G, and H, which is the interference of the programmed selected cell, the fringe field is directed near the channel surface, as shown in Figure 4c; so, it is relatively lower than the interference shown in patterns A, B, C, and D.
In order to alleviate this interference, it is necessary to resolve the insufficient field on a local bias in the read operation. Therefore, in the next section, we describe the proposed method to solve this.

Proposed Read Scheme
To summarize the above, we have compared and analyzed the interference phenomenon for all patterns of one selected cell and two adjacent cells in a scaled-down 3D NAND flash memory. The fundamental reason for the cell-to-cell interference is that the electron concentration changes by the fringe field to the channel surface area near the selected cell.
Therefore, we proposed to alleviate the cell-to-cell interference by controlling the V read bias condition of adjacent cells. Figure 7a-c shows the word line bias schemes 1, 2, and 3 to alleviate the field degradation with each pattern. Figure 7a tries to alleviate the cell-to-cell interference due to electric field degradation by applying V V_P bias, which is higher by ∆V than V read in the case of patterns B and F. Likewise, in Figure 7b, to alleviate the mobility degradation in pattern C and G, as shown in Figure 6b, it is also intended to alleviate the cell-to-cell interference through V V_P bias. However, for the simplification of the scheme, if patterns A and E are read through schemes 1 or 2, the V th shift is minimized through V V_E , which is lower by ∆V than V read , to compensate for the increased electron concentration at channel surface by V V_P . So, schemes 1 and 2 attempt to solve patterns A, B, C, E, F, and G through asymmetric read operation through V V_P and V V_E . For patterns D and H, scheme 3 as shown in Figure 7c is used to alleviate electric field degradation through V V_P bias. However, if the ∆V in scheme 3 is too high, a short channel effect may occur [11]. Therefore, the V V_P bias increase should not be too large.
Schemes 1, 2, and 3 can alleviate the cell-to-cell interference for each pattern, but all schemes have different V th values for ∆V, one ∆V must be set that satisfies for all patterns. Figure 7d shows the V th for each pattern according to ∆V. As mentioned above, for simplicity of the scheme, patterns A and E will perform a read operation using schemes 1 or 2, but the scheme closest to V th of the other patterns will be selected. As a result, in Figure 7d, the point where the patterns A, B, C, and D are gathered is 2 V of ∆V. In pattern A, scheme 1 was closer to the V th values of all patterns than scheme 2. In addition, in Figure 7e, the V th values of all the patterns are adjacent at the point where the ∆V was 1. Furthermore, pattern E is more adjacent to the V th values of all the patterns of Scheme 1 than Scheme 2. Therefore, it is judged that ∆V is compatible with 1.5 in all patterns.  Figure 8 shows the result of the I-V curve characteristic curve of the selected cell for each pattern through TACD simulation using the conventional and proposed read scheme with a ∆V of 1.5 V in a 3D NAND flash memory structure with a gate length of 30 nm and a space length of 10 nm. When comparing Figure 8a,c, the V th shift range was 0.914 V in a conventional read scheme, but it decreased by approximately 67.5% to 0.297 V through the proposed read scheme, and when comparing Figure 8 b,d, the V th shift range was reduced by approximately 70.7% to 0.571 V in the conventional scheme and 0.167 V in the proposed scheme. In addition, it can be seen that the reduction rate of the BL current of the proposed scheme is smaller than that of the conventional scheme. Therefore, the proposed scheme can minimize cell-to-cell interference when the space length is reduced by scaled-down in 3D NAND flash memory.

Conclusions
In this paper, we analyzed the cell-to-cell interference that occurs as the space length decreases. When the adjacent cells were programmed, the V read bias during the read operation affects the channel concentration of the selected cell region due to the insufficient bias condition of the electric field to the channel surface. To solve the cell-to-cell interference caused by the electric field, it is necessary to optimize the bias for adjacent cells, which have the most influence on the channel concentration of the selected cell, in each pattern during the read operation.
Therefore, the proposed read scheme was able to alleviate the V th shift of the selected cells by the program of the adjacent cells by changing the ∆V of 1.5 V in the V read bias applied to the adjacent cells according to each pattern. In a high-stacking structure, such a cell-to-cell interference issue cannot be avoided. Therefore, if the gate and space length are further scaled down and the cell-to-cell interference is severe, through the string patterning and read operation using the proposed scheme, it will be able to alleviate the interference problem.

Conflicts of Interest:
The authors declare no conflict of interest.