An ai-driven 3d nand flash read voltage self-adaptive optimization method and system

Abstract

The application discloses an AI-driven 3D NAND flash memory reading voltage self-adaptive optimization method and system, relates to the technical field of data storage, and comprises the following steps: in the feature perception and prediction stage, multi-dimensional feature vectors of 3D NAND flash memory storage units are collected, input into a lightweight AI prediction model, and a predicted voltage interval, an offset direction probability and a prediction confidence are output; in the adaptive decision and optimization stage, according to the comparison result of the prediction confidence and a preset threshold, a corresponding search strategy is selected, scanning is performed in the predicted voltage interval or around the reference voltage, and an optimal reading voltage is determined; in the hierarchical knowledge construction and real-time fine-tuning stage, multi-source voltage data are acquired and dynamically weighted and fused to obtain an initial voltage, local voltage scanning is performed starting from the initial voltage to complete fine-tuning, and if the reading reliability gain after fine-tuning is insufficient, a gradual safety rollback mechanism is started. The application realizes high-precision accuracy of the reading voltage, and reduces the delay and power consumption.

Description

Technical Field

[0001] This invention relates to the field of data storage technology, specifically to an AI-driven method and system for adaptive optimization of 3D NAND flash memory read voltage. Background Technology

[0002] As the number of stacked layers in 3D NAND flash memory continues to increase, storage density has significantly improved, but this has also brought more severe challenges to data reliability. The threshold voltage (Vth) distribution of memory cells can drift and broaden due to factors such as process variations, programming interference, read interference, data retention characteristic degradation, and cyclic wear. To ensure correct data reading, an accurate read voltage must be applied to distinguish different storage states.

[0003] Traditional read voltage management schemes mainly fall into two categories: The first is using a fixed, factory-preset read voltage. However, traditional fixed voltage schemes rely on a static voltage preset at the factory, which cannot adapt to the dynamic drift of the threshold voltage caused by factors such as wear and temperature during chip use. The fundamental contradiction lies in the mismatch between "static setting" and "dynamic change." This inevitably leads to the read voltage gradually deviating from the optimal position, causing a continuous increase in the original bit error rate, which in turn triggers a series of chain problems such as increased error correction burden, read latency, and increased power consumption, seriously damaging long-term data reliability and system performance. The second approach is real-time voltage calibration (such as read voltage scanning). Real-time full-scan calibration finds the optimal value by traversing the full range of voltages for each read operation. However, this method requires performing dozens to hundreds of trial reads and voltage switching, introducing significant time delays and additional power consumption. The calibration process itself becomes a bottleneck for performance and energy efficiency. This problem is particularly acute in multi-level cells such as QLCs, making this method, while ensuring accuracy, unsuitable for high-performance, low-power memory systems due to its excessive overhead.

[0004] Existing technologies employ several methods to reduce calibration overhead, such as voltage tracking based on error correction code feedback or sampling calibration of local blocks. However, these methods either suffer from slow response times or fail to fully utilize the inherent characteristic interrelationships within 3D NAND flash memory (e.g., different layers, different blocks), resulting in insufficient calibration accuracy and efficiency. This is particularly true in high-density QLC (quad-layer cell) products, where the required number of read voltages is high and the voltage window is narrow, placing even greater demands on calibration techniques.

[0005] Therefore, a technical solution is needed that can intelligently, quickly, and accurately adaptively optimize the read voltage in order to minimize read latency and system overhead while ensuring data reliability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an AI-driven 3D NAND flash memory read voltage adaptive optimization method and system to achieve extremely high calibration accuracy and reduce latency and power consumption.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] First aspect: An AI-driven adaptive optimization method for 3D NAND flash memory read voltage, including:

[0009] In the feature perception and prediction stage, multi-dimensional feature vectors of 3D NAND flash memory cells are collected, input into a lightweight AI prediction model, and the predicted voltage range, offset direction probability, and prediction confidence are output. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features.

[0010] In the adaptive decision-making and optimization phase, based on the comparison between the predicted confidence level and the preset threshold, a corresponding search strategy is selected, and a scan is performed around the predicted voltage range or the reference voltage to determine the optimal reading voltage; the search strategies include a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy;

[0011] In the hierarchical knowledge construction and real-time fine-tuning stage, multi-source voltage data is acquired and dynamically weighted and fused to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is initiated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage.

[0012] Furthermore, the multidimensional feature vector is obtained by concatenating the aging features, environmental features, interference features, topological features, historical features, and process features, and is represented as follows: ,in, The aging characteristic is the effective number of program / erase cycles compensated for by temperature and rate. Environmental characteristics include the mean and variance of current temperature and historical temperature statistics; Interference characteristics include the number of read interferences, the intensity of programming interference, and the data retention time; Topological features are the three-dimensional normalized coordinates of the memory cells within the chip; The historical characteristics are the historical best voltage sequence of this memory cell; The process characteristics are the inherent deviation parameters calibrated for the manufacturing batch.

[0013] Furthermore, the lightweight AI prediction model is a multi-task learning neural network, including a feature embedding layer, a CNN / GRU multi-branch feature extraction layer, and an attention weight modulation layer. The input is the multi-dimensional feature vector, and the output is the predicted voltage range. Offset direction probability And the prediction confidence level C, which ranges from 0 to 1.

[0014] Furthermore, the specific content of the search strategy is as follows: preset a high threshold. and low threshold The output confidence level With the high threshold and low threshold If a comparison is made, Choose a high-confidence, high-precision strategy within the predicted voltage range. Use small, precise steps to sweep inside; if A medium-confidence balancing strategy is selected to predict the voltage range. To expand the scanning range from the center, a variable step size method is used, with denser scanning at the center and sparser scanning at the edges; if A low-confidence robust strategy is selected, falling back to the knowledge base benchmark voltage and performing a wide-range conservative scan around it. Through multi-stage scanning search, the original bit error rate is finally determined. Minimum optimal voltage .

[0015] Furthermore, the acquisition of multi-source voltage data includes the following steps:

[0016] The target physical address is parsed as a multi-level index key, and the voltage values ​​of each level are obtained by querying the hierarchical voltage knowledge base in parallel. and confidence weight The knowledge base reference voltage is obtained through the following weighted fusion calculation. ;

[0017] A list of addresses surrounding the target cell is generated based on a predefined proximity strategy. After batch querying the latest voltage values ​​corresponding to the addresses in the list and removing outliers, the arithmetic mean of the remaining voltage values ​​is calculated to obtain the average voltage of the spatially adjacent cells. ;

[0018] Read the N most recent historical best voltage sequences of the target cell The time-series predicted voltage is obtained through an exponential smoothing model. Smoothing factor Dynamically adjusted based on sequence volatility For model state;

[0019] Collect 5-7 key features to form a simplified feature vector, use a dedicated neural network or decision tree model with less than 1KB of input parameters, and complete inference and output AI-predicted voltage within 1 microsecond. .

[0020] Furthermore, the hierarchical voltage knowledge base adopts a multi-level architecture, including chip-level, die-level, block / layer-level, and page / word-line-level, with each level of voltage data associated with a corresponding confidence weight.

[0021] Furthermore, the initial voltage is obtained by dynamically weighting and fusing the multi-source voltage data, and is expressed as:

[0022] ;

[0023] in The dynamic weighting coefficients are allocated in real time based on the long-term historical accuracy of each voltage source and its feature similarity to the current scene.

[0024] Furthermore, if the read reliability gain after fine-tuning is insufficient, a gradual safety fallback mechanism will be initiated, specifically including:

[0025] The original bit error rate RBER_new obtained using the final read voltage V_final after fine-tuning is compared with the estimated bit error rate RBER_base using the reference voltage V_base. The relative gain G = (RBER_base - RBER_new) / RBER_base is calculated. If the relative gain G is lower than the preset threshold, or the original bit error rate RBER_new exceeds the error correction tolerance, it is determined that the gain is insufficient, and the gradual safety back-off mechanism is triggered.

[0026] The progressive safety rollback mechanism includes:

[0027] Expanding the window and rescanning: With the final reading voltage V_final as the center, expand the search window by 3 to 5 times, and rescan with a step size of 2 to 3 voltage levels to find a better voltage;

[0028] The higher-level reference voltage rescan step is used as a reference. If the scan reading fails in the expanded window rescan step, the statistical reference voltage of the previous level is queried and used for direct reading.

[0029] Streamline the global scan steps: If the rescan step using a higher-level reference voltage still fails, perform a non-uniform step size scan over the entire voltage range [V_min, V_max]. In the high-probability voltage distribution region identified by historical data, set a smaller step size of 1 voltage level for fine searching. In the low-probability edge region of the distribution, set a larger step size of 4-5 voltage levels for rapid cross-search to find the available voltage.

[0030] Furthermore, it also includes the model tuning and evolution stage, specifically including:

[0031] Continuously collect end-to-end data from the feature perception and prediction stage, the adaptive decision-making and optimization stage, and the hierarchical knowledge construction and real-time fine-tuning stage;

[0032] The parameters of the lightweight AI prediction model are fine-tuned in real time based on the collected data to perform online incremental learning;

[0033] The lightweight AI prediction model is periodically retrained using the collected big data, and the parameters are optimized for offline deep optimization.

[0034] Based on the performance statistics, the search strategy is adaptively modified and evolved by optimizing the strategy threshold.

[0035] The second aspect: an AI-driven 3D NAND flash memory read voltage adaptive optimization system, comprising:

[0036] The feature perception and prediction module is used to collect multi-dimensional feature vectors of 3D NAND flash memory cells, input them into a lightweight AI prediction model, and output the predicted voltage range, offset direction probability, and prediction confidence. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features.

[0037] An adaptive decision-making and optimization module is used to select a corresponding search strategy based on the comparison result between the predicted confidence level and a preset threshold, and to perform a scan around the predicted voltage range or the reference voltage to determine the optimal reading voltage; the search strategy includes a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy;

[0038] The hierarchical knowledge construction and real-time fine-tuning module is used to acquire multi-source voltage data and perform dynamic weighted fusion to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is initiated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage.

[0039] The beneficial effects of this invention are:

[0040] This invention utilizes a lightweight AI model to accurately predict narrow ranges of voltage drift based on multi-dimensional state features, significantly reducing the initial search range and substantially lowering overhead from the source. By constructing a hierarchical reference voltage tree and performing rapid page-level fine-tuning, a balance between global accuracy and real-time response is achieved. Model optimization and evolution ensure that the system can be continuously optimized throughout the chip's lifecycle. Therefore, this invention significantly reduces latency and power consumption while maintaining extremely high calibration accuracy, comprehensively improving the reliability, performance, and energy efficiency of the storage system. Attached Figure Description

[0041] Figure 1 A flowchart illustrating an AI-driven adaptive optimization method for 3D NAND flash memory read voltage;

[0042] Figure 2 This is a schematic diagram of an AI-driven 3D NAND flash memory read voltage adaptive optimization system. Detailed Implementation

[0043] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] See Figures 1-2 The present invention provides a technical solution: Example

[0045] An AI-driven adaptive optimization method for 3D NAND flash memory read voltage, such as Figure 1 As shown, it includes:

[0046] In the feature perception and prediction stage, multi-dimensional feature vectors of 3D NAND flash memory cells are collected, input into a lightweight AI prediction model, and the predicted voltage range, offset direction probability, and prediction confidence are output. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features.

[0047] The multidimensional feature vector is obtained by concatenating the aging feature, environmental feature, interference feature, topological feature, historical feature, and process feature, and is represented as follows: ,in, The aging characteristic is the effective number of program / erase cycles compensated for by temperature and rate. Environmental characteristics include the mean and variance of current temperature and historical temperature statistics; Interference characteristics include the number of read interferences, the intensity of programming interference, and the data retention time; Topological features are the three-dimensional normalized coordinates of the memory cells within the chip; The historical characteristics are the historical best voltage sequence of this memory cell; The process characteristics are the inherent deviation parameters calibrated for the manufacturing batch.

[0048] The lightweight AI prediction model is a multi-task learning neural network, including a feature embedding layer, a CNN / GRU multi-branch feature extraction layer, and an attention weight modulation layer. The input is the multi-dimensional feature vector, and the output is the predicted voltage range. Offset direction probability And the prediction confidence level C, which ranges from 0 to 1.

[0049] Optionally, the lightweight AI prediction model can be a high-performance machine learning model such as gradient boosting decision tree or random forest, or a Bayesian neural network deployed at the edge to output uncertainty (confidence) more naturally. It can also achieve the mapping from features to prediction intervals and provide confidence estimates.

[0050] By leveraging multi-dimensional feature perception and AI-driven proactive prediction, "preventive calibration" is achieved. This proactively adjusts the voltage before errors occur, eliminating the response lag defects of traditional solutions, significantly improving data reliability, and preventing read failures. Experiments show that the AI ​​prediction model can reduce the initial search range by an average of over 70%. Combined with a confidence-level adaptive strategy, it avoids invalid fine scanning and redundant wide-range scanning, reducing single-order optimization latency from hundreds of microseconds to tens of microseconds, significantly reducing the number of trial reads, and simultaneously lowering system power consumption.

[0051] In the adaptive decision-making and optimization phase, based on the comparison between the predicted confidence level and the preset threshold, a corresponding search strategy is selected to perform a scan within the predicted voltage range or around the reference voltage to determine the optimal reading voltage. The search strategies include a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy.

[0052] The specific content of the search strategy is as follows: preset a high threshold. and low threshold The output confidence level With the high threshold and low threshold If a comparison is made, Choose a high-confidence, high-precision strategy within the predicted voltage range. Use small, precise steps to sweep inside; if A medium-confidence balancing strategy is selected to predict the voltage range. To expand the scanning range from the center, a variable step size method is used, with denser scanning at the center and sparser scanning at the edges; if A low-confidence robust strategy is selected, falling back to the knowledge base benchmark voltage and performing a wide-range conservative scan around it. Through multi-stage scanning search, the original bit error rate is finally determined. Minimum optimal voltage .

[0053] The confidence-based adaptive strategy decision-making mechanism innovatively uses the uncertainty (confidence level) of AI prediction as the core basis for strategy selection, achieving precise matching between resource allocation and optimization needs.

[0054] In the hierarchical knowledge construction and real-time fine-tuning stage, multi-source voltage data is acquired and dynamically weighted and fused to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is initiated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage.

[0055] In this embodiment, acquiring multi-source voltage data includes the following steps:

[0056] The target physical address is parsed as a multi-level index key, and the voltage values ​​of each level are obtained by querying the hierarchical voltage knowledge base in parallel. and confidence weight The knowledge base reference voltage is obtained through the following weighted fusion calculation. The hierarchical voltage knowledge base adopts a multi-level architecture, including chip level, die level, block / layer level, and page / word line level, with each level of voltage data associated with a corresponding confidence weight.

[0057] A list of addresses surrounding the target cell is generated based on a predefined proximity strategy. After batch querying the latest voltage values ​​corresponding to the addresses in the list and removing outliers, the arithmetic mean of the remaining voltage values ​​is calculated to obtain the average voltage of the spatially adjacent cells. .

[0058] Read the N most recent historical best voltage sequences of the target cell The time-series predicted voltage is obtained through an exponential smoothing model. Smoothing factor Dynamically adjusted based on sequence volatility This represents the model state.

[0059] Collect 5-7 key features to form a simplified feature vector, use a dedicated neural network or decision tree model with less than 1KB of input parameters, and complete inference and output AI-predicted voltage within 1 microsecond. .

[0060] In another alternative embodiment, the hierarchical voltage knowledge base can be replaced with a graph neural network-managed structure, where nodes represent storage units at different levels and edges represent spatial / logical proximity relationships. By using graph propagation algorithms to update and query voltage knowledge, the complex spatial correlations in the storage array can be captured and utilized more flexibly.

[0061] In this embodiment, the initial voltage is obtained by dynamically weighting and fusing the multi-source voltage data, and is expressed as:

[0062] ;

[0063] in The dynamic weighting coefficients are allocated in real time based on the long-term historical accuracy of each voltage source and its feature similarity to the current scene.

[0064] In this embodiment, the step-by-step safety fallback mechanism, which is activated if the read reliability gain after fine-tuning is insufficient, specifically includes:

[0065] The original bit error rate RBER_new obtained using the final read voltage V_final after fine-tuning is compared with the estimated bit error rate RBER_base using the reference voltage V_base. The relative gain G = (RBER_base - RBER_new) / RBER_base is calculated. If the relative gain G is lower than the preset threshold (e.g., 5%), or the original bit error rate RBER_new exceeds the error correction tolerance, it is determined that the gain is insufficient, triggering the gradual safety back-off mechanism.

[0066] The progressive safety rollback mechanism includes:

[0067] Expanding the window and rescanning: With the final reading voltage V_final as the center, expand the search window by 3 to 5 times, and rescan with a step size of 2 to 3 voltage levels to find a better voltage;

[0068] The higher-level reference voltage rescan step is used as a reference. If the scan reading fails in the expanded window rescan step, the statistical reference voltage of the previous level is queried and used for direct reading.

[0069] Streamline the global scan steps: If the rescan step using a higher-level reference voltage still fails, perform a non-uniform step size scan over the entire voltage range [V_min, V_max]. In the high-probability voltage distribution region identified by historical data, set a smaller step size of 1 voltage level for fine searching. In the low-probability edge region of the distribution, set a larger step size of 4-5 voltage levels for rapid cross-search to find the available voltage.

[0070] The hierarchical knowledge base and real-time fine-tuning dual-layer architecture distinguish between deep optimization in the background and instantaneous fine-tuning in the foreground, ensuring both global optimization accuracy and meeting the real-time response requirements of instantaneous reading, and adapting to the narrow voltage window requirements of high-density flash memory products such as QLC.

[0071] Furthermore, the method of the present invention also includes a model tuning and evolution stage, specifically including:

[0072] Continuously collect end-to-end data from the feature perception and prediction stage, the adaptive decision-making and optimization stage, and the hierarchical knowledge construction and real-time fine-tuning stage;

[0073] The parameters of the lightweight AI prediction model are fine-tuned in real time based on the collected data to perform online incremental learning;

[0074] The lightweight AI prediction model is periodically retrained using the collected big data, and the parameters are optimized for offline deep optimization.

[0075] Based on the performance statistics, the search strategy is adaptively modified and evolved by optimizing the strategy threshold.

[0076] It employs a deep multi-task learning model to model complex nonlinear relationships. During the model tuning and evolution stages, it continuously trains the model and adjusts the strategy with new data. The prediction accuracy improves with usage time, and the optimization capability continues to iterate as the chip ages, achieving "the more you use it, the more accurate it becomes," which can effectively combat the performance degradation at the end of the life cycle.

[0077] Example 2: An AI-driven adaptive optimization system for 3D NAND flash memory read voltage, such as Figure 2 As shown, it includes:

[0078] The feature perception and prediction module is used to collect multi-dimensional feature vectors of 3D NAND flash memory cells, input them into a lightweight AI prediction model, and output the predicted voltage range, offset direction probability, and prediction confidence. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features.

[0079] The adaptive decision-making and optimization module is used to select a corresponding search strategy based on the comparison result between the predicted confidence level and the preset threshold, and to perform scanning in the predicted voltage range or around the reference voltage to determine the optimal reading voltage; the search strategy includes a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy.

[0080] The hierarchical knowledge construction and real-time fine-tuning module is used to acquire multi-source voltage data and perform dynamic weighted fusion to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is initiated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage.

[0081] It also includes a model tuning and evolution module, specifically used for:

[0082] Continuously collect end-to-end data from the feature perception and prediction stage, the adaptive decision-making and optimization stage, and the hierarchical knowledge construction and real-time fine-tuning stage;

[0083] The parameters of the lightweight AI prediction model are fine-tuned in real time based on the collected data to perform online incremental learning;

[0084] The lightweight AI prediction model is periodically retrained using the collected big data, and the parameters are optimized for offline deep optimization.

[0085] Based on the performance statistics, the search strategy is adaptively modified and evolved by optimizing the strategy threshold.

[0086] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An AI-driven adaptive optimization method for 3D NAND flash memory read voltage, characterized in that, include: In the feature perception and prediction stage, multi-dimensional feature vectors of 3D NAND flash memory cells are collected, input into a lightweight AI prediction model, and the predicted voltage range, offset direction probability, and prediction confidence are output. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features. In the adaptive decision-making and optimization phase, based on the comparison between the predicted confidence level and the preset threshold, a corresponding search strategy is selected, and a scan is performed around the predicted voltage range or the reference voltage to determine the optimal reading voltage; the search strategies include a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy; In the hierarchical knowledge construction and real-time fine-tuning stage, multi-source voltage data is acquired and dynamically weighted and fused to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning and obtain the final read voltage. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is activated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage. The specific content of the search strategy is as follows: preset a high threshold. and low threshold The output confidence level With the high threshold and low threshold If a comparison is made, Choose a high-confidence, high-precision strategy within the predicted voltage range. Use small, precise steps to sweep inside; if A medium-confidence balancing strategy is selected to predict the voltage range. To expand the scanning range from the center, a variable step size method is used, with denser scanning at the center and sparser scanning at the edges; if A low-confidence robust strategy is selected, falling back to the knowledge base benchmark voltage and performing a wide-range conservative scan around it. Through multi-stage scanning search, the original bit error rate is finally determined. Minimum optimal voltage ; The acquisition of multi-source voltage data includes the following steps: The target physical address is parsed as a multi-level index key, and the voltage values ​​of each level are obtained by querying the hierarchical voltage knowledge base in parallel. and confidence weight The knowledge base reference voltage is obtained through the following weighted fusion calculation. ; A list of addresses surrounding the target cell is generated based on a predefined proximity strategy. After batch querying the latest voltage values ​​corresponding to the addresses in the list and removing outliers, the arithmetic mean of the remaining voltage values ​​is calculated to obtain the average voltage of the spatially adjacent cells. ; Read the N most recent historical best voltage sequences of the target cell The time-series predicted voltage is obtained through an exponential smoothing model. Smoothing factor Dynamically adjusted based on sequence volatility For model state; Collect 5-7 key features to form a simplified feature vector, use a dedicated neural network or decision tree model with less than 1KB of input parameters, and complete inference and output AI-predicted voltage within 1 microsecond. ; The initial voltage is obtained by dynamically weighting and fusing the multi-source voltage data, and is expressed as follows: ； in The dynamic weighting coefficients are allocated in real time based on the long-term historical accuracy of each voltage source and its feature similarity to the current scene. If the read reliability gain after fine-tuning is insufficient, a gradual safety fallback mechanism will be initiated, specifically including: The original bit error rate RBER_new obtained using the final read voltage V_final after fine-tuning is compared with the estimated bit error rate RBER_base using the reference voltage V_base. The relative gain G = (RBER_base - RBER_new) / RBER_base is calculated. If the relative gain G is lower than the preset threshold, or the original bit error rate RBER_new exceeds the error correction tolerance, it is determined that the gain is insufficient, and the gradual safety back-off mechanism is triggered. The progressive safety rollback mechanism includes: Expanding the window and rescanning: With the final reading voltage V_final as the center, expand the search window by 3 to 5 times, and rescan with a step size of 2 to 3 voltage levels to find a better voltage; The higher-level reference voltage rescan step is used as a reference. If the scan reading fails in the expanded window rescan step, the statistical reference voltage of the previous level is queried and used for direct reading. Streamline the global scan steps: If the rescan step using a higher-level reference voltage still fails, perform a non-uniform step size scan over the entire voltage range [V_min, V_max]. In the high-probability voltage distribution region identified by historical data, set a smaller step size of 1 voltage level for fine searching. In the low-probability edge region of the distribution, set a larger step size of 4-5 voltage levels for rapid cross-search to find the available voltage.

2. The AI-driven adaptive optimization method for 3D NAND flash memory read voltage according to claim 1, characterized in that, The multidimensional feature vector is obtained by concatenating the aging feature, environmental feature, interference feature, topological feature, historical feature, and process feature, and is represented as follows: ,in, The aging characteristic is the effective number of program / erase cycles compensated for by temperature and rate. Environmental characteristics include the mean and variance of current temperature and historical temperature statistics; Interference characteristics include the number of read interferences, the intensity of programming interference, and the data retention time; Topological features are the three-dimensional normalized coordinates of the memory cells within the chip; The historical characteristics are the historical best voltage sequence of this memory cell; The process characteristics are the inherent deviation parameters calibrated for the manufacturing batch.

3. The AI-driven adaptive optimization method for 3D NAND flash memory read voltage according to claim 1, characterized in that: The lightweight AI prediction model is a multi-task learning neural network. It includes a feature embedding layer, a CNN / GRU multi-branch feature extraction layer, and an attention weight modulation layer. The input is the multidimensional feature vector, and the output is the predicted voltage range. Offset direction probability And the prediction confidence level C, which ranges from 0 to 1.

4. The AI-driven adaptive optimization method for 3D NAND flash memory read voltage according to claim 1, characterized in that: The hierarchical voltage knowledge base adopts a multi-level architecture, including chip level, die level, block / layer level, and page / word line level, with each level of voltage data associated with a corresponding confidence weight.

5. The AI-driven adaptive optimization method for 3D NAND flash memory read voltage according to claim 1, characterized in that: It also includes the model tuning and evolution stage, specifically including: Continuously collect end-to-end data from the feature perception and prediction stage, the adaptive decision-making and optimization stage, and the hierarchical knowledge construction and real-time fine-tuning stage; The parameters of the lightweight AI prediction model are fine-tuned in real time based on the collected data to perform online incremental learning; The lightweight AI prediction model is periodically retrained using the collected big data, and the parameters are optimized for offline deep optimization. Based on the performance statistics, the search strategy is adaptively modified and evolved by optimizing the strategy threshold.

6. An AI-driven adaptive optimization system for 3D NAND flash memory read voltage, characterized in that: include: The feature perception and prediction module is used to collect multi-dimensional feature vectors of 3D NAND flash memory cells, input them into a lightweight AI prediction model, and output the predicted voltage range, offset direction probability, and prediction confidence. The multi-dimensional feature vectors include aging features, environmental features, interference features, topological features, historical features, and process features. An adaptive decision-making and optimization module is used to select a corresponding search strategy based on the comparison result between the predicted confidence level and a preset threshold, and to perform a scan around the predicted voltage range or the reference voltage to determine the optimal reading voltage; the search strategy includes a high-confidence accurate strategy, a medium-confidence balanced strategy, and a low-confidence robust strategy; The specific content of the search strategy is as follows: preset a high threshold. and low threshold The output confidence level With the high threshold and low threshold If a comparison is made, Choose a high-confidence, high-precision strategy within the predicted voltage range. Use small, precise steps to sweep inside; if A medium-confidence balancing strategy is selected to predict the voltage range. To expand the scanning range from the center, a variable step size method is used, with denser scanning at the center and sparser scanning at the edges; if A low-confidence robust strategy is selected, falling back to the knowledge base benchmark voltage and performing a wide-range conservative scan around it. Through multi-stage scanning search, the original bit error rate is finally determined. Minimum optimal voltage ; The hierarchical knowledge construction and real-time fine-tuning module is used to acquire multi-source voltage data and perform dynamic weighted fusion to obtain an initial voltage. Starting from the initial voltage, a local voltage scan is performed to complete the fine-tuning. If the read reliability gain is insufficient after fine-tuning, a gradual safety fallback mechanism is activated. The multi-source voltage data includes the knowledge base reference voltage, the average voltage of spatially adjacent units, the time-series predicted voltage, and the AI ​​predicted voltage. The acquisition of multi-source voltage data includes the following steps: The target physical address is parsed as a multi-level index key, and the voltage values ​​of each level are obtained by querying the hierarchical voltage knowledge base in parallel. and confidence weight The knowledge base reference voltage is obtained through the following weighted fusion calculation. ; A list of addresses surrounding the target cell is generated based on a predefined proximity strategy. After batch querying the latest voltage values ​​corresponding to the addresses in the list and removing outliers, the arithmetic mean of the remaining voltage values ​​is calculated to obtain the average voltage of the spatially adjacent cells. ; Read the N most recent historical best voltage sequences of the target cell The time-series predicted voltage is obtained through an exponential smoothing model. Smoothing factor Dynamically adjusted based on sequence volatility For model state; Collect 5-7 key features to form a simplified feature vector, use a dedicated neural network or decision tree model with less than 1KB of input parameters, and complete inference and output AI-predicted voltage within 1 microsecond. ; The initial voltage is obtained by dynamically weighting and fusing the multi-source voltage data, and is expressed as follows: ； in The dynamic weighting coefficients are allocated in real time based on the long-term historical accuracy of each voltage source and its feature similarity to the current scene. If the read reliability gain after fine-tuning is insufficient, a gradual safety fallback mechanism will be initiated, specifically including: The original bit error rate RBER_new obtained using the final read voltage V_final after fine-tuning is compared with the estimated bit error rate RBER_base using the reference voltage V_base. The relative gain G = (RBER_base - RBER_new) / RBER_base is calculated. If the relative gain G is lower than the preset threshold, or the original bit error rate RBER_new exceeds the error correction tolerance, it is determined that the gain is insufficient, and the gradual safety back-off mechanism is triggered. The progressive safety rollback mechanism includes: Expanding the window and rescanning: With the final reading voltage V_final as the center, expand the search window by 3 to 5 times, and rescan with a step size of 2 to 3 voltage levels to find a better voltage; The higher-level reference voltage rescan step is used as a reference. If the scan reading fails in the expanded window rescan step, the statistical reference voltage of the previous level is queried and used for direct reading. Streamline the global scan steps: If the rescan step using a higher-level reference voltage still fails, perform a non-uniform step size scan over the entire voltage range [V_min, V_max]. In the high-probability voltage distribution region identified by historical data, set a smaller step size of 1 voltage level for fine searching. In the low-probability edge region of the distribution, set a larger step size of 4-5 voltage levels for rapid cross-search to find the available voltage.

Images