KR-20260065497-A - Learning Model System for Reflecting Characteristics of ReRAM and Operating Method Thereof

KR20260065497AKR 20260065497 AKR20260065497 AKR 20260065497AKR-20260065497-A

Abstract

A learning model system that reflects the characteristics of a resistance change memory and a method of operation thereof are presented. The learning method that reflects the characteristics of a resistance change memory proposed in the present invention includes the steps of receiving characteristic values for predicting current-voltage curves under multiple conditions through an input unit, performing LSTM (Long Short-Term Memory)-based learning through a prediction unit to predict current-voltage curves under multiple conditions, and outputting a next prediction result based on the previous learning result for the prediction of current-voltage curves under multiple conditions through an output unit.

Inventors

정성엽
이태헌
조민선
이오준
정진우

Assignees

고려대학교 세종산학협력단

Dates

Publication Date: 20260508
Application Date: 20250917
Priority Date: 20241101

Claims (10)

A step of receiving characteristic values for predicting current-voltage curves under multiple conditions through an input unit; A step of predicting current-voltage curves under multiple conditions by performing time-series-based artificial neural network learning through a prediction unit; and A step of outputting the next prediction result based on the previous learning result for current-voltage curve prediction under multiple conditions through an output unit. A method of operation of a learning model that reflects the simulation of the characteristics of a resistance change memory including
In paragraph 1, The step of receiving characteristic values for predicting current-voltage curves under multiple conditions through the above-mentioned input unit is: To predict current-voltage curves under multiple conditions, input characteristic values including voltage, predetermined variables affecting the learning model results, time, rate of change of voltage, and past current. Operation method of a learning model reflecting the characteristics of a resistance change memory.
In paragraph 1, The step of predicting current-voltage curves under multiple conditions by performing LSTM-based learning through the prediction unit described above is: Input variables regarding sequential states are given for characteristic values including voltage, predetermined variables affecting the learning model results, time, voltage change rate, and past current, respectively. Operation method of a learning model reflecting the characteristics of a resistance change memory.
In paragraph 3, To predict current-voltage curves under multiple conditions, training for LSTM-based current-voltage curve prediction is performed using input variables regarding each of the sequential states. Operation method of a learning model reflecting the characteristics of a resistance change memory.
In paragraph 1, The step of outputting the next prediction result based on the previous learning result for current-voltage curve prediction under multiple conditions through the above output unit is: Predicting the current value for the voltage value by using the LSTM-based learning result from the above prediction unit, and using the field vector for the previous learning result as input for the next learning. Operation method of a learning model reflecting the characteristics of a resistance change memory.
An input unit that receives characteristic values for predicting current-voltage curves under multiple conditions; A prediction unit that predicts current-voltage curves under multiple conditions by performing LSTM (Long Short-Term Memory) based learning; and An output unit that outputs the next prediction result based on the previous learning result for current-voltage curve prediction under multiple conditions A learning model system that reflects the simulation of the characteristics of a resistance change memory including
In paragraph 6, The above input unit is, To predict current-voltage curves under multiple conditions, input characteristic values including voltage, predetermined variables affecting the learning model results, time, rate of change of voltage, and past current. A learning model system that simulates the characteristics of resistance change memory.
In paragraph 6, The above prediction unit is, Input variables regarding sequential states are given for characteristic values including voltage, predetermined variables affecting the learning model results, time, voltage change rate, and past current, respectively. A learning model system that simulates the characteristics of resistance change memory.
In paragraph 8, The above prediction unit is, To predict current-voltage curves under multiple conditions, training for LSTM-based current-voltage curve prediction is performed using input variables regarding each of the sequential states. A learning model system that simulates the characteristics of resistance change memory.
In paragraph 6, The step of outputting the next prediction result based on the previous learning result for the current-voltage curve prediction under the above plurality of conditions is, Predicting the current value for the voltage value by using the LSTM-based learning result from the above prediction unit, and using the field vector for the previous learning result as input for the next learning. A learning model system that simulates the characteristics of resistance change memory.

Description

Learning Model System for Reflecting Characteristics of ReRAM and Operating Method Thereof The present invention relates to a learning model system that reflects the characteristics of Resistance Random Access Memory (ReRAM) and a method of operation thereof. Resistive switching memory devices, which utilize the characteristic of a resistive switching material whose resistance value changes depending on voltage conditions, can operate with only a simple structure of electrode layer-resistive switching material layer-electrode layer and have advantages in terms of fast switching speed and low-voltage operation. Unlike other next-generation non-volatile memory devices, these resistive switching memory devices can increase integration density through a simple process, and the manufacturing cost is low. Resistive change memory devices are generally implemented in the form of a crossbar array, and a selector is placed in each memory cell to operate the device so that current can flow only to the desired cell. A resistance switching memory device comprises a material layer having two states, wherein a conductive region with a low resistance value (Low Resistance Stage; LRS) is formed depending on voltage conditions, and a state with a high resistance value (High Resistance Stage; HRS) is formed when the conductive region is broken, and it has a switching function through the distinction between the two states. Based on this, since single-layer resistive switching memory devices had a problem of insufficient switching capability because they could not have sufficiently large resistance in the HRS state, a double-layer material was applied to the resistive switching memory device to secure switching capability while having sufficiently large resistance in the HRS state. However, unlike single-layer resistive memory devices, it is unclear at which point in the dual layer filament formation and extinction caused by oxygen vacancies occur in dual-layer resistive memory devices. Consequently, dual-layer resistive memory devices require inspection when high-density integrated circuits are fabricated and operated, and it is difficult to analyze the cause of malfunctions or operational failures. Resistive switching memory is a next-generation AI semiconductor that performs computations within memory, going beyond the von Neumann architecture of existing first-generation AI. It is expected to increase power efficiency and improve computational performance, and is becoming increasingly important as a technology that will enable the application of ICT technologies. Meanwhile, in resistance change memory devices composed of a metal-insulator-metal structure, various switching characteristics occur depending on the type of insulator or the interface state between the metal and the insulator; therefore, there are various models that simulate the related physical variables and device characteristics. Most existing methods for calculating model variables are limited to presenting maximum and minimum parameter ranges and manually searching for variables using restricted algorithms. In particular, this approach has limitations in terms of computational efficiency and the reflection of generalized characteristics. Furthermore, commercially available circuit simulation-based optimization techniques are limited to modeling transistors or diodes and are not suitable for devices requiring time-domain analysis techniques, such as Resistive Random Access Memory (ReRAM). Accordingly, the present invention proposes a method that enables the simulation of non-linear I-V that varies depending on the application of various input waveforms and can improve computational efficiency through a technique for calculating the range of model variables. Figure 1 is a diagram showing the structure of a SiOx-based resistance change memory. Figure 2 is a diagram illustrating the state of a resistance change memory. FIG. 3 is a flowchart illustrating a learning method that reflects the characteristics of a resistance change memory according to an embodiment of the present invention. FIG. 4 is a diagram illustrating an LSTM structure according to an embodiment of the present invention. FIG. 5 is a diagram showing the configuration of a learning model system that reflects the characteristics of a resistance change memory according to an embodiment of the present invention. FIG. 6 is a diagram showing the results of visualizing a learning model according to one embodiment of the present invention. FIG. 7 is a diagram showing the accuracy results of a learning model according to one embodiment of the present invention. Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. Figure 1 is a diagram showing the structure of a SiOx-based resistance change memory. Referring to FIG. 1, a SiOx-based resistive switching memory corresponding to the resistive switching memory (ReRAM) mentioned herein may be im