# US 12,112,813 B2 · Device for determining read reference voltage of a memory device and operating method thereof

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,112,813 B2
- Document ID: US12112813B2
- Status: Granted
- Kind: B2
- Application number: 17/847,056
- Filed: 2022-06-22
- Published: 2023-09-21
- Granted: 2024-10-08
- Earliest claimed: 2022-01-28
- Assignees: SK hynix Inc.; Korea Advanced Institute of Science and Technology
- Applicants: SK hynix Inc.; Korea Advanced Institute of Science and Technology
- Inventors: Sunyoung Jo; Jungwuk Park; Younghyun Park; Sang Ho Yun; Jaekyun Moon
- Web page: https://patentgrep.com/patent/US12112813B2

## CPC

- G11C 16/3431: {Circuits or methods to detect disturbed nonvolatile memory cells, e.g. which still read as programmed but with threshold less than the program verify threshold or read as erased but with threshold greater than the erase verify threshold, and to reverse the disturbance via a refreshing programming or erasing step}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G11C 11/54: using elements simulating biological cells, e.g. neuron
- G11C 11/5642: {Sensing or reading circuits; Data output circuits}
- G11C 16/26: Sensing or reading circuits; Data output circuits
- G11C 16/28: using differential sensing or reference cells, e.g. dummy cells
- G11C 29/021: {in voltage or current generators}
- G11C 29/028: {with adaption or trimming of parameters}
- G11C 29/50004: {of threshold voltage}
- G11C 16/0483: {comprising cells having several storage transistors connected in series}
- G11C 16/3404: {Convergence or correction of memory cell threshold voltages; Repair or recovery of overerased or overprogrammed cells}
- G11C: STATIC STORES (semiconductor memory devices H10B)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A device includes a threshold voltage distribution estimation network configured to generate an estimated distribution using a feature distribution and read trial information, a set of feature distributions generated from a plurality of threshold voltage distributions for a plurality of pages of a memory device, and a read reference voltage estimation network configured to generate a read reference voltage from the estimated distribution. The read trial information includes a read trial vector and an output value, the output value being generated by applying the read trial vector to a threshold voltage distribution for a page to be read among the plurality of threshold voltage distributions.

## Claims

### Claim 1

A device comprising:
a threshold voltage distribution estimation network configured to generate an estimated distribution matrix using a feature distribution matrix and read trial information, the threshold voltage distribution estimation network including a feature distribution generating circuit and a combination circuit, the feature distribution generating circuit configured to generate the feature distribution matrix from a plurality of threshold voltage distributions for a plurality of pages of a memory device, the combination circuit configured to generate the estimation distribution matrix by combining a plurality of vectors included in the feature distribution matrix using a combination coefficient vector; and
a read reference voltage estimation network configured to generate a read reference voltage from the estimated distribution matrix,
wherein the read trial information includes a read trial vector and an output value, the output value being generated by applying the read trial vector to a threshold voltage distribution for a page to be read among the plurality of threshold voltage distributions, and
wherein the threshold voltage distribution estimation network further includes a coefficient estimation network configured to generate the combination coefficient vector from the read trial information.

### Claim 2 (depends on 1)

The device of claim 1, wherein the coefficient estimation network includes a main neural network generating the combination coefficient vector according to the read trial information.

### Claim 3 (depends on 2)

The device of claim 2, wherein the coefficient estimation network further includes an input neural network receiving the read trial information and providing an output to the main neural network, and an output neural network receiving an output of the main neural network and providing the combination coefficient vector.

### Claim 4 (depends on 1)

The device of claim 1, wherein the feature distribution matrix corresponds to a basis matrix extracted from a matrix including the plurality of threshold voltage distributions for the plurality of pages using a Non-negative Matrix Factorization (NMF) technique.

### Claim 5 (depends on 1)

The device of claim 1, wherein the read reference voltage estimation network includes a first neural network configured to receive the estimated distribution matrix and generate the read reference voltage.

### Claim 6 (depends on 5)

The device of claim 5, wherein the read reference voltage estimation network further includes a second neural network receiving an output of the first neural network and generating the read reference voltage.

### Claim 7 (depends on 1)

The device of claim 1, wherein the memory device stores multibit data including a most significant bit (MSB) and a least significant bit (LSB) in a memory cell, and
wherein the read trial vector includes a plurality of elements selected according to a read trial, the read trial being performed for the MSB or the LSB.

### Claim 8 (depends on 7)

The device of claim 7, wherein a number of the plurality of elements included in the read trial vector is determined by a pattern of the MSB or a pattern of the LSB, each of the pattern of the MSB and the pattern of the LSB depending on data levels stored in the memory cell.

### Claim 9 (depends on 7)

The device of claim 7, wherein the output value is generated using an integration vector and the threshold voltage distribution for the page to be read, the integration vector being determined according to the read trial vector.

### Claim 10 (depends on 1)

The device of claim 1, wherein the read reference voltage is a read reference voltage vector including a plurality of read reference voltages.

### Claim 11

A method for determining a read reference voltage distribution, the method comprising:
generating read trial information corresponding to a threshold voltage distribution for a page among a plurality of threshold voltage distributions for a plurality of pages;
generating a combination coefficient vector by applying the read trial information to a coefficient estimation network;
generating, using a feature distribution generating circuit, a feature distribution matrix from the plurality of threshold voltage distributions for the plurality of pages;
generating, using a combination circuit, an estimated distribution matrix by combining a plurality of vectors included in the feature distribution matrix using the combination coefficient vector; and
estimating a read reference voltage by applying the estimated distribution matrix to a read reference voltage estimation network,
wherein generating the feature distribution matrix includes:
generating a threshold voltage distribution matrix from the plurality of threshold voltage distributions for the plurality of pages; and
generating a basis matrix corresponding to the feature distribution matrix by applying a specific technique to the threshold voltage distribution matrix.

### Claim 12 (depends on 11)

The method of claim 11, generating the read trial information and generating the combination coefficient vector are performed multiple times.

### Claim 13 (depends on 11)

The method of claim 11,
wherein the basis matrix is generated by applying a Non-negative Matrix Factorization (NMF) technique to the threshold voltage distribution matrix.

### Claim 14 (depends on 11)

The method of claim 11, further comprising a learning operation of the coefficient estimation network, the learning operation including:
generating a plurality of pieces of read trial information for a page among the plurality of pages;
calculating a cumulative loss by accumulating loss function values when the plurality of pieces of read trial information are applied to the coefficient estimation network; and
updating weights of the coefficient estimation network in a direction to minimize the cumulative loss.

### Claim 15 (depends on 11)

The method of claim 11, further comprising a learning operation of the read reference voltage estimation network, the learning operation including:
calculating a loss function value according to the read reference voltage generated when the estimated distribution matrix is applied to the read reference voltage estimation network; and
updating weights of the read reference voltage estimation network in a direction to minimize the loss function value.

### Claim 16 (depends on 11)

The method of claim 11, wherein estimating the read reference voltage includes:
applying the estimated distribution matrix to a convolutional neural network (CNN) to generate an output; and
generating the read reference voltage based on the output of the CNN.

### Claim 17 (depends on 11)

The method of claim 11, wherein the read reference voltage is a read reference voltage vector including a plurality of read reference voltages.