A temperature drift self-compensated neuron circuit based on volatile threshold transition memristor
By using a temperature-drift self-compensating neuron circuit based on a volatile threshold-switching memristor and modulating the input magnitude with an on-chip thermal sensor, the problem of output frequency drift of the threshold-switching memristor under temperature changes is solved, thus achieving stability of the neuron circuit and low-energy neuromorphic computation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PEKING UNIV
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-19
AI Technical Summary
In existing neuromorphic computing systems, the device parameters of threshold switching memristors have significant temperature dependence characteristics, which leads to a significant impact on the performance of the device and neuron circuits when the ambient temperature changes, and there is a lack of efficient temperature drift compensation methods.
A temperature drift self-compensating neuron circuit based on a volatile threshold switching memristor was designed. The input magnitude of the neuron is modulated by a homogeneous on-chip thermal sensor, and the charging current of the integrated-discharge circuit is controlled by the temperature drift compensation circuit to achieve stability compensation of the output pulse frequency.
The neuron output pulse frequency remains relatively stable when the temperature changes, reducing hardware overhead and power consumption. It is suitable for application scenarios with significant temperature changes, such as training/inference scenarios for large models and cloud/edge applications.
Smart Images
Figure CN122242599A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial neural network technology, specifically relating to a temperature drift self-compensating neuron circuit based on a volatile threshold switching memristor for input encoding and signal processing in a spiking neural network. Background Technology
[0002] In recent years, with the continuous advancement of artificial intelligence technologies, exemplified by large language models, the computational demands for large model training and inference have been increasing daily. Traditional computer systems employ the von Neumann architecture, where there is a physical separation between storage and processing units. During operation, a large amount of data needs to be moved between these units, severely impacting speed and power consumption, thus limiting further improvements in computing power. Neuromorphic computing systems, inspired by the human brain, are considered a strong candidate for overcoming the "von Neumann bottleneck," possessing advantages such as high parallelism, event-driven characteristics, low latency, and high energy efficiency, demonstrating broad application prospects.
[0003] Current neuromorphic computing systems generally encode information into pulse sequences of different frequencies for processing, with both encoding and processing relying on neuronal circuits. Threshold-switching memristors, due to their device dynamics similar to biological neurons, are widely used in leaky-integrate-and-fire (LIF) neuronal circuits, offering significant advantages in area and power consumption compared to traditional CMOS-based designs. However, the device parameters of threshold-switching memristors (such as threshold voltage, holding voltage, and on / off resistance) exhibit significant temperature dependence. Changes in ambient temperature cause these parameters to alter significantly, further impacting the performance of the device, neuronal circuits, and even the neural network. Since the operating scenarios and environments of neuromorphic computing systems are unpredictable, robustness to operate stably at varying temperatures is crucial, necessitating appropriate compensation strategies. While software post-processing can implement relevant compensation algorithms, it undoubtedly introduces additional power consumption and latency. To date, no research reports have been published on how to construct efficient temperature-drift-compensated neurons with minimal hardware overhead, a simple feedback circuit, and compatibility with threshold-switching devices. Summary of the Invention
[0004] To overcome the current lack of efficient temperature drift compensation neurons, this invention provides a temperature drift self-compensating neuron circuit for signal processing in neuromorphic computing systems based on a volatile threshold switching memristor. It can modulate the input size of neurons through homogeneous on-chip thermal sensors, reducing the temperature drift of neuron firing frequency under the same input when the temperature changes, thereby improving the robustness of neurons and neuromorphic computing systems.
[0005] The technical solution provided by this invention is as follows:
[0006] A temperature-drift self-compensating neuron circuit based on a volatile threshold-switching memristor includes an input section, an integrated-discharge circuit based on the volatile threshold-switching memristor, and a temperature-drift compensation circuit. The input section can be a current source or a voltage source connected in series with a resistor. The integrated-discharge circuit based on the volatile threshold-switching memristor includes a volatile threshold-switching memristor TS, a load resistor R1 connected in series with the memristor TS, and a parallel capacitor C1. One end of the memristor TS is connected to C1, and the other end is connected in series with R1 and then grounded. The node where the memristor TS is connected to C1 is the membrane potential V. mem The node connected to R1 is the output pulse node; the temperature drift compensation circuit includes a current source, a voltage source, a temperature sensor (similar to a volatile threshold switching memristor), and an NMOS transistor M1. The two ends of the temperature sensor are connected to the voltage source V1 and the current source I1, respectively. The node V1 connected to I1 is... ctrl The gate of transistor M1 is connected to control the gate voltage of transistor M1, and the drain of transistor M1 is connected to V. mem Node, source grounded.
[0007] Furthermore, the volatile threshold switching memristor is a Mott memristor based on an insulator-metal phase change material or an OTS memristor based on a chalcogenide. Its device parameters have temperature-dependent characteristics, decreasing with increasing temperature and increasing with decreasing temperature within a certain range.
[0008] Furthermore, the capacitor C1 of the integrated-emission circuit accumulates input excitation. When its voltage reaches the threshold voltage of the memristor TS, the memristor TS undergoes a threshold transition, releasing a pulse. Through the above process, the input current or voltage signal is converted into a pulse sequence output. The temperature sensor of the temperature drift compensation circuit can sense changes in ambient temperature and reflect these changes in the gate voltage of transistor M1. By adjusting the charging current of the integrated-emission circuit through transistor M1, the charging time of capacitor C1 is modulated, further controlling the output pulse frequency of the integrated-emission circuit.
[0009] Furthermore, as the ambient temperature rises, the threshold voltage and off-state resistance of the threshold switching memristor TS decrease, and the resistance of the temperature sensor in the temperature drift self-compensation circuit decreases, leading to V ctrl As the voltage drop across the node increases, the gate voltage of transistor M1 increases, leading to an increase in the current flowing through transistor M1, which in turn causes V to... mem The voltage drop at the node decreases, causing the output pulse frequency of the output node to decrease, returning to the frequency level at lower temperatures; when the ambient temperature drops, the resistance of the temperature sensor in the temperature drift self-compensation circuit increases, leading to V ctrlThe voltage drop at the node decreases, the gate voltage of transistor M1 decreases, the current through transistor M1 decreases, resulting in V... mem The voltage drop across the node increases, which in turn increases the output pulse frequency of the output node, bringing it closer to the frequency level at higher temperatures. This compensates for the temperature drift effect of the integrated-output circuit's output frequency.
[0010] The beneficial effects of this invention are as follows:
[0011] This invention provides a temperature-drift self-compensating neuron circuit based on a volatile threshold-switching memristor. Utilizing the temperature-dependent parameter changes of the threshold-switching memristor, a temperature sensor homogeneous with the memristor is constructed. By sensing temperature changes, the charging / discharging behavior of the integration-discharge circuit is modulated, thus maintaining a relatively stable output pulse frequency despite temperature variations. This invention offers an effective compensation scheme for the temperature drift effect of LIF neurons. It eliminates the need for complex software post-processing, achieving temperature drift compensation directly through on-chip hardware circuitry. It is suitable for applications with significant temperature variations, such as large-scale model training / inference scenarios, and cloud / edge-side applications. This facilitates larger-scale integration and lower power consumption, demonstrating broad development prospects. Attached Figure Description
[0012] Figure 1 This is a circuit diagram of a temperature drift self-compensating neuron for signal processing in neuromorphic computing based on a volatile threshold switching memristor, as described in this invention. R1 / R2, C1, and M1 represent resistor, capacitor, and transistor, respectively. in For the input current, V in V is the input voltage, V1 is a constant voltage source, I1 is a constant current source, and V mem The membrane potential is V, and the output is the output pulse voltage. ctrl The transistor gate control voltage is used, and the input is a current source (I). in ) or a voltage source with a series resistor (V in and R2).
[0013] Figure 2 This is a flowchart illustrating the operation of the temperature drift self-compensating neuron circuit based on a volatile threshold switching memristor for signal processing in neuromorphic computing, according to the present invention.
[0014] Figure 3 For a conventional LIF neuron circuit based on a volatile threshold-switching memristor under different input I... in The output pulse frequency at temperature.
[0015] Figure 4 This invention relates to a temperature drift self-compensating neuron circuit based on a volatile threshold-switching memristor under different input I... in The output pulse frequency at temperature.
[0016] Figure 5 To address the temperature drift self-compensating neuron circuit based on a volatile threshold switching memristor in this invention, the input I is adjusted under different ambient temperatures. in The output pulse waveforms are the same. Detailed Implementation
[0017] The present invention will be further clearly and completely described below with reference to the accompanying drawings and specific embodiments.
[0018] This invention proposes a temperature drift self-compensating neuron circuit based on a volatile threshold switching memristor, which can compensate for the temperature drift effect of LIF neurons when the temperature changes with minimal hardware overhead. Figure 1 This is a circuit diagram of a temperature drift self-compensating neuron based on a volatile threshold-switching memristor, as described in this invention. The neuron circuit includes an input section, an integration-discharge circuit based on the volatile threshold-switching memristor, and a temperature drift compensation circuit. The input section can be a current source or a voltage source connected in series with a resistor. The integration-discharge circuit based on the volatile threshold-switching memristor includes a volatile threshold-switching memristor TS, a load resistor R1 connected in series with the memristor TS, and a parallel capacitor C1. One end of the memristor TS is connected to C1, and the other end is connected in series with R1 and then grounded. The node where the memristor TS is connected to C1 is the membrane potential V. mem The node connected to R1 is the output pulse node; the temperature drift compensation circuit includes a current source I1, a voltage source V1, a temperature sensor homogeneous with the volatile threshold switching memristor TS, and an NMOS transistor M1. The two ends of the temperature sensor are connected to the voltage source V1 and the current source I1, respectively. The node V1 connected to I1 is... ctrl The gate of transistor M1 is connected to control its gate voltage, and the drain of transistor M1 is connected to the V of the integrated discharge circuit. mem Node, source grounded.
[0019] Figure 2 This is a flowchart illustrating the operation of the temperature drift self-compensating neuron based on a volatile threshold switching memristor according to the present invention. Specifically, when the ambient temperature rises, the resistance of the temperature sensor in the temperature drift self-compensating circuit decreases, leading to V... ctrl As the voltage drop across the node increases, the gate voltage of transistor M1 increases. This increased gate voltage leads to a decrease in the channel resistance of transistor M1, V mem The voltage drop at the node decreases, meaning the charging voltage of the integrated-discharge circuit decreases, slowing down the charging process of C1. This, in turn, leads to a decrease in the output pulse frequency of the output node in the integrated-discharge circuit, maintaining the output frequency at a level close to that before the temperature change. When the ambient temperature drops, the resistance of the temperature sensor in the temperature drift self-compensation circuit increases, causing V... ctrlThe voltage drop at the node decreases, the gate voltage of transistor M1 decreases, the current through transistor M1 decreases, resulting in V... mem The voltage drop across the node increases, which in turn increases the output pulse frequency of the output node, bringing it closer to the frequency level at higher temperatures. This compensates for the temperature drift effect of the integrated-output circuit's output frequency.
[0020] The volatile threshold-switching memristor described in this invention can be a Mott memristor based on an insulator-metal phase change material, or an OTS memristor based on chalcogenides. Further, the Mott memristor can be based on VO2 or NbO. x Devices made of materials such as Se and Te, OTS memristors can be devices based on chalcogen elements such as Se and Te.
[0021] The volatile threshold-switching memristor, while possessing volatility and threshold-switching characteristics, also exhibits temperature-dependent device parameters. Specifically, key parameters such as the device's off-state resistance, threshold voltage, and holding voltage change with temperature. Within a certain range, these parameters decrease as temperature increases and increase as temperature decreases.
[0022] Furthermore, taking the VO2 MOT memristor as an example, when the ambient temperature rises, the threshold voltage and off-state resistance of the threshold switching memristor TS decrease. Therefore, without a temperature drift self-compensation circuit, when the input I... in or V in At the same time, the higher the ambient temperature, the higher the pulse frequency output by the output node, resulting in temperature drift. Figure 3 Without temperature drift compensation circuitry, ordinary LIF neuron circuits operate under different I... in The output frequency varies with input and different ambient temperatures. When input I... in At the same time, the output frequency increases with the rise in ambient temperature, exhibiting a severe temperature drift phenomenon.
[0023] Figure 4 This invention relates to a temperature drift self-compensating neuron circuit based on a VO2 volatile threshold switching memristor, which adapts to varying ambient temperatures (298~318 K) and different input I values. in The output pulse frequency variation curve is shown below. Figure 5 This invention relates to a temperature drift self-compensating neuron circuit based on a VO2 volatile threshold switching memristor, which adjusts for temperature drift under different ambient temperatures when the input I... in The output voltage waveform is shown at 600μA. It can be seen that when the ambient temperature changes, the voltage at the same input I... in Under these conditions, the neuron output frequency did not show a significant shift, compared to Figure 3The severe temperature drift phenomenon that occurs in ordinary LIF neuron circuits is addressed by the temperature drift self-compensating neuron circuit based on volatile threshold switching memristors, which can greatly reduce the temperature drift of neuron output pulse frequency when the ambient temperature changes.
[0024] This invention proposes a novel temperature-drift self-compensating neuron circuit. It utilizes a temperature sensor, homogeneous with a volatile threshold-switching memristor, to regulate the charging voltage of the integrated-fire circuit via a transistor. This further modulates the temperature drift effect of the output pulse frequency caused by temperature changes in the threshold-switching device parameters within the integrated-fire circuit, allowing the output pulse frequency of the integrated-fire circuit to remain relatively stable despite ambient temperature variations. This invention achieves self-compensation for temperature drift through on-chip hardware, eliminating the need for complex software post-processing. It is suitable for applications with significant temperature variations, such as large-scale model training / inference scenarios and cloud / edge-side applications. It facilitates larger-scale integration and lower power consumption, demonstrating broad development prospects.
[0025] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.
Claims
1. A temperature-drift self-compensating neuron circuit based on a volatile threshold-switching memristor, characterized in that, It includes an input section, an integrated-discharge circuit based on a volatile threshold switching memristor, and a temperature drift compensation circuit; the input section is a current source or a voltage source connected in series with a resistor; The integrated-discharge circuit based on a volatile threshold-switching memristor includes a volatile threshold-switching memristor TS, a load resistor R1 connected in series with the memristor TS, and a parallel capacitor C1. One end of the memristor TS is connected to C1, and the other end is connected in series with R1 and then grounded. The node where the memristor TS and C1 are connected is at the film potential V. mem The node connected to R1 is the output pulse node; the temperature drift compensation circuit includes a current source, a voltage source, a temperature sensor (similar to a volatile threshold switching memristor), and an NMOS transistor M1. The two ends of the temperature sensor are connected to the voltage source V1 and the current source I1, respectively. The node V1 connected to I1 is... ctrl The gate of transistor M1 is connected to control the gate voltage of transistor M1, and the drain of transistor M1 is connected to V. mem Node, source grounded.
2. The temperature drift self-compensating neuron circuit as described in claim 1, characterized in that, The volatile threshold switching memristor is a Mott memristor based on an insulator-metal phase change material or an OTS memristor based on a chalcogenide. Its device parameters have temperature-dependent characteristics, decreasing with increasing temperature and increasing with decreasing temperature within a certain range.
3. The temperature drift self-compensating neuron circuit as described in claim 1, characterized in that, The capacitor C1 of the integrated-emission circuit accumulates input excitation. When its voltage reaches the threshold voltage of the memristor TS, the memristor TS undergoes a threshold transition, releasing a pulse. Through this process, the input current or voltage signal is converted into a pulse sequence output. The temperature sensor of the temperature drift compensation circuit senses changes in ambient temperature and reflects these changes in the gate voltage of transistor M1. Transistor M1 adjusts the charging current of the integrated-emission circuit, thereby modulating the charging time of capacitor C1 and further controlling the output pulse frequency of the integrated-emission circuit.
4. The temperature drift self-compensating neuron circuit as described in claim 1, characterized in that, As the ambient temperature rises, the threshold voltage and off-state resistance of the threshold switching memristor TS decrease, and the resistance of the temperature sensor in the temperature drift self-compensation circuit decreases, leading to V ctrl As the voltage drop across the node increases, the gate voltage of transistor M1 increases, leading to an increase in the current flowing through transistor M1, which in turn causes V to... mem The voltage drop at the node decreases, causing the output pulse frequency of the output node to decrease, returning to the frequency level at lower temperatures; when the ambient temperature drops, the resistance of the temperature sensor in the temperature drift self-compensation circuit increases, leading to V ctrl The voltage drop at the node decreases, the gate voltage of transistor M1 decreases, the current through transistor M1 decreases, resulting in V... mem The voltage drop across the node increases, which in turn increases the output pulse frequency of the output node, bringing it closer to the frequency level at higher temperatures. This compensates for the temperature drift effect of the integrated-output circuit's output frequency.