1. Introduction
Due to the characteristics of high storage density and low bit cost, 3D NAND flash has become the mainstream structure of high-performance memory chips [
1,
2]. Meanwhile, sensor systems are becoming increasingly complex in scale and function, and it is critical to be able to collect, store, and manage large amounts of data. The nodes of sensor networks are capable of sensing physical quantities such as temperature, humidity, and pressure in the environment. These data, after being processed and stored, can be used to monitor and control various applications such as smart homes, healthcare, industrial automation, and more [
3]. Therefore, ensuring complete and accurate access to stored data is critical to the stable operation of sensor systems. NAND flash memory, as a non-volatile memory device, is the most widely used data storage method in modern sensor systems due to its fast access [
4,
5]. At the same time, NAND flash memory has high endurance and shock resistance, which makes it possible to store data stably for a long time under extreme operating conditions [
6]. In addition, advanced sensor systems, such as wireless sensor networks (WSNs), use an approach whereby they sense, store, merge, and send to monitoring [
7]. Sensors store generated data in flash memory cells and perform in-network calculations when required [
8,
9,
10]. Advanced flash memory technology provides sensor networks with powerful capabilities for the collection, storage, and processing of data.
To meet the demand for large data capacity growth, 3D NAND flash keeps scaling and storage layers are constantly stacked [
11,
12]. As a result, some reliability problems, such as data disturbance and retention, become more serious and neighbor wordline interference (NWI) is one of these problems [
13,
14]. NWI refers to the effect wherein the programmed cell causes the threshold voltage (
) of its neighbor cells to shift [
15]. The
shift can result in an incorrect state read from the victim cell [
16]. NWI has long existed in high-density flash memory and is difficult to completely eliminate. Different from floating gate-based 2D NAND, due to changes such as 3D vertical architecture, poly-silicon channels, and charge traps, the NWI effect in 3D NAND needs further investigation [
17,
18,
19].
Previous studies on 3D NAND NWI were undertaken mainly from the perspective of trapped electrons, suggesting that the charge trap layer (CTL) between adjacent wordlines (WLs) was programmed into redundant electrons, shifting the
of neighbor cells [
20]. Therefore, some studies adopted methods such as asymmetric
[
21] and the CTL blocking process [
22] to eliminate the NWI that resulted from the electron reduction between WLs. Unfortunately, as the process size of WL and isolation becomes smaller, these redundant electrons become hard to avoid and eliminate. Furthermore, it is almost impossible to accurately define and measure the influence scopes of trapped electrons. In fact, NWI interference comes from potential interference. Based on the channel potential continuity theory in 3D NAND structure, the trapped electrons of other flash cells can affect the channel potential of the victim cell and then interfere with its read
[
23]. From the perspective of channel potential, the NWI can be investigated more intuitively and accurately. In this way, some studies reduced NWI by enhancing the
applied to neighbor cells [
24] or by adopting reverse order program [
25].
However, the NWI, especially in 3D NAND, is affected by many electronic factors, and there is no unified approach to deal with various ramifications of NWI. Research has shown that the electrons between the WL space contribute about 40% of NWI [
26], so there remain other disturbing factors. In fact, in 3D NAND, the device operating parameters of read and write will also affect NWI, and it should be accurately integrated in a physical model. Moreover, the NWI varies greatly under different combinations of cell states, inevitably resulting in various degrees of residual NWI.
Therefore, in this paper, a new 3D NAND device model combined with channel potential was established to further investigate the NWI mechanism. It integrates the effects of various device parameters, such as , , , and different on NWI. An NWI analysis scheme by channel potential is proposed to optimize the original method by trapped electron. Using technology computer-aided design (TCAD) simulation, a 3D NAND structure model with eight-layer WLs was built. The physical device model indicated that the NWI effect is closely related to device operating bias. TCAD simulations also confirmed this and found that the applied device bias can result in a significant difference in channel potential changes caused by NWI. Thus, the theory of channel potential superposition and the local drain-induced barrier lowering (DIBL) effect were utilized to describe the mechanism of NWI generation, providing an efficient way to quickly analyze NWI problems from a device perspective. On this basis, an adaptive bitline voltage () countermeasure for 3D NAND arrays was proposed to reduce the NWI on triple-level cells (TLC) cells rationally and maximally. Finally, these proposed theories and methods were successfully verified by TCAD experiments and 3D NAND chip tests. This scheme in 3D NAND provides a flexible and extensible solution to alleviate NWI, promising to efficiently enhance data reliability.
2. Simulation Set-Up for NWI
Figure 1a shows the bit cost scalable (BiCS) architecture of a 3D NAND flash, where wordline (WL) and isolation are alternately stacked to form multiple layers for storage. A large number of memory holes vertically cross the layers and form individual memory cells.
Figure 1b illustrates the structure of a metal-oxide–nitride-oxide–silicon (MONOS) memory cell used in the TCAD simulation. From outside to inside, it consists of a tungsten WL, block layer (BLK), charge trap layer (CTL), tunnel oxide layer (TNL), poly-silicon channel, and dielectric core. The BLK consists of barrier metal/block high-k/block oxide and the TNL consists of oxide/nitride/oxide (O/N/O), both of which are three-layer structures.
Figure 2a shows the TCAD modeling structure of a single 3D NAND bitline (BL) string, and a single MONOS memory cell in the red box. The device doping concentrations are shown as color levels.
Figure 2b–d shows the composition and material details of a MONOS cell, a TNL, and a BLK respectively. The simulation assumes rotational symmetry along the vertical central axis of the memory hole, so that the cylindrical mathematical method can be used to simulate a 3D cylindrical device with a 2D mesh. The string parameters of the TCAD model are shown in
Figure 2e.
and
are ON pitch length, and
is radius of core.
,
, and
are thickness of channel, O/N/O, and CTL, respectively.
/
and
/
are n-type and p-type doping concentrations of drain and source, respectively.
and
are electron and hole trap density respectively, and
is electron trap energy in CTL [
27].
Three-dimensional NAND cells are programmed forward from sourceline (SL) to bitline (BL) (low WL1 to high WL6). WL0 and WL7 serve as dummy WLs that do not participate in the program, designed to mitigate the potential difference between SGS/SGD and WLs. The simulation uses an incremental single-pulse program voltage , of which the high-period increases from to . The NAND test program is executed with the incremental step pulse program (ISPP) voltage . The cell for simulation and test is of the TLC type, and each cell contains eight states (three bits/cell). Different program times can ensure that the memory cell is programmed to the specific according to the required state. The bitline voltage is 0.5V in both program and read, where is also equal to . The value is read according to the WL gate voltage when is 10nA, which can be obtained from the - curve in simulation. The NWI effect on WL4 caused by WL5 was simulated. The difference of WL4 read before and after WL5 program is the shift of WL4 caused by NWI, which will be explained next.
3. Physical Device Model of NWI
The NWI effect is caused by the programming of neighbor flash cells, resulting in interference to the victim cell. In practice, NWI is calculated as (shift) = (NWI) − (initial). This is the common approach used to measure the NWI value in both TCAD simulation and 3D NAND chip test. On the process, first program WLn−1, and then program WLn. At this time, read the of WLn as (initial). On this basis, program WLn+1 and finally read the of victim WLn as (NWI).
In this study, an NWI model of a 3D NAND device was established according to the NWI generation environment and the theoretical expression at the device level was derived. In the linear and saturation regions of the standard
-
curve, the drain current
of a flash cell can be modeled by the gate-source voltage
and drain-source voltage
as:
where
is the charge-carrier effective mobility,
is the gate oxide capacitance per unit area,
W is the gate width, and
L is the gate length.
When WLn is read for the first time, the bias voltage applied to the device is shown in
Figure 3a. It is worth mentioning that the actual 3D NAND device performs read operations under the all bit line (ABL) sense method [
28]. As the cell current of the sense amplifier circuit reaches
, the
of the read cell is approximately equal to or slightly larger than the cell
. Thus, the read cell is in the saturation region of the
-
curve, and the (
,
) point falls near the advance pinch-off point. At this point, there is:
According to (2), the drain current is modeled as (3). Thus, in this case, the sense current in the ABL sense is obtained as (4):
where
The value of
can be calculated as (7), where
is a constant for specific device materials and structures. It is controlled by the characteristic length
as (8), based on the short channel effect model:
where
is the equivalent oxide thickness of the BLK, CTL, and TNL gate stacks,
is the thickness of channel, and
and
are the dielectric constants of Si and gate oxide.
After programming WLn+1, we read WLn a second time. The bias voltage applied to the device is shown in
Figure 3b. In this case, the WLn cell is still in the saturation region of the
-
curve. The gate voltage is
and the drain-source voltage is now reduced to
. Therefore,
can be written as:
The newly trapped electrons of WLn+1 enhance the local negative potential. Due to the same
, a channel potential drop
is actually generated. The WLn+1 cell is in the linear region with a small
. Therefore, the
of WLn+1 can be written as (11). The WLn+1 channel is in the pass state under
, so the potential drop
is small and the
can be ignored as (12):
All the newly trapped electrons between WLn and WLn+1 can be regarded as a parasitic cell, which will affect the
of WLn. Since the channel potential drop (
) of the parasitic cell is also very small, it is in the linear region of the
-
curve. Therefore, the
and
in (13) can be ignored. The channel conduction parameter of the parasitic cell is considered to be
similar to
, so
of the parasitic cell is derived as (14):
After solving the equations, including (4), (6), (9), (10), (12), and (14), the
change of WLn between two read operations is obtained. Namely, the NWI measurement result is obtained as follows:
The proposed theoretical device model in (16) shows the main factors leading to the NWI effect, including , , , and . These factors are adjustable parameters that are worth investigating for NWI problems. Briefly, according to these factors, the NWI value can be reduced by increasing , decreasing the amount of parasitic cell electrons, decreasing ( reflects the program level of WLn), decreasing (lower program state of WLn+1), increasing channel length (weakening DIBL effect), and so on. These provide valuable theoretical guidance and mitigation schemes for the subsequent research. Moreover, there is the question of how exactly these factors work and induce NWI during device operation. To explore this question, the channel potential mechanism for the NWI effect is further investigated in the next section.
6. Discussion and Analysis
In this section, we present the discussion about the proposed NWI physical device model and adaptive scheme, in terms of further analysis and future research directions.
6.1. NWI Physical Device Model and Channel Potential Analysis
In this study, a multi-parameter physical device model is proposed based on the actual performance of NWI in 3D NAND flash. The model is used to reflect the influence of each parameter on NWI and their interaction. Among these,, , , , and the material/structure parameters of the device are the main influencing factors of NWI. The model shows the compositional structure of NWI and provides an explanation of each parameter, which is essential for further investigation of the NWI mechanism. , as the channel current of the read cell, directly reflects the threshold voltage magnitude of WLn. Contrary to the previous view that NWI is mainly caused by redundant electrons, this suggests that the storage state of the affected cell plays a crucial role in determining the magnitude of NWI. Within the appropriate range, a slightly lower and a higher storage state are beneficial for reducing NWI. as read pass voltage, especially applied on WLn+1 and WLn-1, has a significant effect on the NWI of WLn. In the channel potential analysis of NWI, it was found that can directly compensate the potential loss of WLn due to the potential superposition and the local DIBL effect caused by the short channel. The large potential gradient near WLn puts it in a state where it is susceptible to interference from the neighbor potential variations. The difference of applied on WLn+1 between the two reads of WLn can be increased while satisfying the normal conduction of the channel under all other WL cells, so that the channel potential barrier of WLn is at the same level prior to and after all the upper cells are programmed. Through chip tests and parameter training, properly raising the adjacent can avoid causing derived program disturbances and read disturbances. The bitline voltage is used as a virtual drain voltage in the WLn read phase, and the voltage can directly affect the data read results through the transfer of the passed channel. However, this transfer will be inevitably weakened by all of the upper programmed cells on the drain side; however, this loss is only slight due to the presence of the read pass voltage. This weakening is most pronounced only near WLn+1, as at the location of it is not possible to fully conduct the channel. is defined as the equivalent threshold voltage of electrons between cells, which represents the fraction of electrons in the wordline isolation layer caused by diffusion or drift, that still have a contribution to the threshold voltage of adjacent cells. This fraction of the electrons is difficult to avoid, but its number can be reduced by some novel program methods to cover the effects or by material and process blocking. is the state in which the WLn+1 cell is programmed and also one of the reasons why WLn suffers from differential NWI. The structural parameters of the device mostly depend on the device process and technology nodes. In particular, the channel length is critical for the generation of NWI effect. As the device size continues to scale, further reductions in the cell channel length may cause even more cells to have a large NWI on WLn, exacerbating the data shift. Materials for flash memory devices have more research options and by improving the process or by using advanced dielectric and charge trapping materials, the diffusion and drift of trapped electrons in the CTL can be reduced, with lower charge loss and NWI.
The channel potential analysis method based on the potential superposition and local DIBL effect was described in detail in
Section 4. This section serves as an extension of the physical device model at the channel potential level, since it can be seen that the key parameters in the physical device model mostly related to the channel potential. The physical device model adopts a more intuitive description method for the convenience of the test developer’s quick judgment. The channel potential analysis serves as the mechanism for NWI generation. It uses the local DIBL effect to essentially explain that the source of the NWI is the virtual drain potential loss in the adjacent channel. On the other hand, both the physical device model and the channel potential analysis provide a theoretical basis for the subsequently proposed adaptive
scheme, which guides
to modulate the channel potential and thus reduce NWI.
In terms of benefits, the NWI physical device model is the first physical model to reflect the NWI effect in BiCS 3D NAND flash memory, which is one of the latest technologies in industrial 3D NAND flash devices. In fact, the doping structure of 2D NAND devices is very different from the virtual source/drain structure of 3D NAND. Therefore, the proposed physical model is an update based on the latest 3D NAND devices. In addition, the structure of the NWI physical device model is composed of the actual EPR parameters, which is more convenient for reflecting the parameter influence on the NWI results. Thus, the model serves as a better reference in the application of real chip operations. With more adjustable electrical parameters, this model can help device testers and designers to evaluate NWI noise faster and more accurately in the field. Most importantly, the NWI physical device model highlights the use of , , , , and , providing theoretical support not only for the adaptive scheme, but also for many other optimization strategies. Of course, in terms of possible drawbacks, this model is an ideal model based on the standard MONOS flash cell and is mostly used for the qualitative analysis of NWI patterns. An accurate quantitative analysis of NWI results by the ideal model is not the best choice. In practice, NWI results are more often obtained by NAND platform testing and then backwards extrapolation. In addition, the model is proposed for NWI problems, so its application to other reliability problems such as data retention or process defect interference also remains to be investigated.
In the practical chip test, different patterns can be adopted according to different application scenarios and performance requirements. After the basic parameter training is completed, new suitable compensation parameters can be obtained by testing the corresponding pattern for the NWI environment. The identification of influential adjustable parameters and cell states is essential for further exploration of the NWI mechanism. The NWI physical device model integrates various relevant device factors, reflecting the interaction of each factor and its contribution to NWI. It also provides theoretical support for the development of appropriate optimization schemes. The exploration of potential factors and novel structures for NWI will continue.
6.2. Analysis of Adaptive Vbl Countermeasure
In this paper, a scheme utilizing adaptive bitline voltage regulation for NWI in the read phase is proposed and verified by 3D NAND chip test. The scheme is supported by the physical device model and the channel potential theory. The channel potential superposition and the local DIBL effect caused by NWI will lead to a decrease in the virtual drain potential of the read cell. The bitline voltage can be transferred to the virtual drain of the read cell through the passed channel on the drain side. Therefore, the virtual drain voltage loss can be basically compensated for by a series of adaptive values in the read phase. This scheme has a small increase in , and is only implemented in the read phase after all programming is completed, with minimal impact on other operations. This scheme requires a quick pre-read of the programmed WLn+1 in the read phase, which is used to determine the state of WLn+1.
In fact, as shown in
Figure 10b, the NWI differences between some adjacent state combinations are not significant. In such cases, within the tolerance range, the same
can be applied to these regions, which can reduce the frequency of voltage changes and improve the efficiency of this scheme. Therefore, for pre-reading WLn+1, the states with similar
compensation values can be grouped into the same zone. If only A to D, E to F, and G states are considered as three zones, two read voltage levels can be used to distinguish between the low, medium, and high states of WLn+1. If the grouping is reduced to two zones, only one read voltage level is required, which can shorten the time to achieve fast reading. Similarly, for reading WLn, the eight states of TLC can also be divided into several zones based on the NWI proximity. For instance, if the A to D and E to G states are divided into high and low zones, then in a normal read verify phase, only two adaptive
values need to be used in the corresponding high (A–D) and low (E–G) verification zones. During a single normal read verification process, only two adaptive
values need to be used in the corresponding high and low state verification intervals. A reasonable partition can greatly improve the implementation efficiency of this scheme.
On the other hand, the NWI variations caused by the stacking of upper wordlines can be overcome by the cooperation of and . In today’s 3D NAND flash technology, the stacking of wordline layers (>200 layers) may lead to a non-negligible accumulation of channel resistance after all cells have been programmed. To address this issue without introducing additional noise, appropriate compensation for can be achieved by subsequent pattern tests.
In addition, even if the random combination of upper programmed WLs still causes channel resistance variations among different strings (which can be understood as differences in virtual drain potential loss of WLn), the adaptive scheme can also cover this effect. This is because the pre-read result of WLn+1 in the adaptive scheme can actually reflect the channel resistance difference. If the virtual drain potential of WLn shows string differences due to the stacking of upper WLs, the pre-read of WLn+1 will also show differences. Depending on the result of WLn+1, the compensation will also change, even jumping to another zone to cover this effect. The optimization accuracy of such difference depends on the fault tolerance of the NAND flash memory system. As mentioned above, this scheme can minimize such interference with a balance between efficiency and compensation accuracy.
The study focuses on exploring strategies to improve NWI by using adaptive . Physical device model and channel potential analysis also provide directions, such as / electrical parameter adjustment schemes, novel program schemes based on cell state combinations, and advanced device structures and materials. These studies will help us further investigate the set of NWI issues and provide more reliable data storage performance for NAND flash-based sensor systems.
7. Conclusions
In this study, to alleviate data disturbances and NWI effects in a 3D NAND flash, a physical device model was established and an adaptive countermeasure was proposed, with the NWI mechanism analysis based on channel potential. First, in TCAD, it was found that the channel potential change under read bias condition could reflect the actual NWI more accurately than the floating case. Thus, an NWI device model composed of flash operation parameters was built, which could intuitively suggest the contribution of , , and concerned to NWI. NWI is composed of two main processes: channel potential superposition and local DIBL effect. Then, it was demonstrated that the local DIBL effect is the main cause of NWI by TCAD simulations with wordline gaps. Finally, an adaptive countermeasure was proposed to mitigate NWI in 3D NAND arrays. In this way, the is boosted in stages according to the combinations of WLn and WLn+1 states, which in turn significantly balances and minimizes the NWI of TLC cells. These patterns were successfully verified in both TCAD and 3D NAND chip test. This study will help to advance NWI optimization and further support efficient prediction of data reliability in 3D NAND flash memory-based sensor systems.