Reserve pool computing network based on temperature regulation and control nonvolatile access and application of reserve pool computing network

Abstract

The invention discloses a reserve pool computing network based on temperature regulation and control nonvolatile access. The reserve pool computing network comprises a room temperature reserve layer module, a temperature interface module and a low temperature reading layer module which are arranged in sequence, the room temperature storage layer module realizes short-term memory by utilizing the dynamic characteristics of a nonvolatile memory and is used for quickly processing signals; the temperature interface module is used for realizing nonlinear mapping of signal characteristics between a room temperature environment and a low temperature environment, realizing smooth information connection across temperature domains and supporting end-to-end gradient transmission; and the low-temperature reading layer module is used for realizing long-term memory by utilizing the stability characteristic of a nonvolatile memory and is used for long-term data storage and weight updating. The reservoir computing network is used for image recognition, signal classification, speech processing and other artificial intelligence computing scenarios. According to the invention, short-term dynamic response is realized in a room temperature region, long-term stable storage is realized in a low temperature region, short-term memory and long-term memory functions are considered, and the precision, energy efficiency and reliability of the neural network are effectively improved.

Description

Technical Field

[0001] The present invention relates to temperature characteristic control of non-volatile memory (NVM) and is used for constructing a hybrid temperature reservoir computing network, belonging to the technical field of neuromorphic computing (reservoir computing) and non-volatile memory devices. Background Art

[0002] Traditional von Neumann architectures face a bottleneck in the separation of memory and computation when handling artificial intelligence tasks, resulting in high power consumption and significant latency. Neuromorphic computing, with its brain-inspired architecture, parallel processing, and energy efficiency, has become a promising candidate for emerging computing. Reservoir computing (RC) networks, with their advantages of simple training and low energy consumption, have become a promising candidate for achieving high-performance computing. Their hardware implementation requires both short-term memory (STM) and long-term memory (LTM) capabilities. At the hardware implementation level, neuromorphic devices based on non-volatile memory (NVM) exhibit unique advantages. While existing NVM-based approaches for constructing RC networks (such as the ion dynamics of memristors, polarization switching of FeFETs, and ferroelectric polarization of HZO-based capacitors) can achieve both STM and LTM, they still face room for improvement in stability, scalability, and energy efficiency due to device non-idealities, operational pulse complexity, and speed-integration trade-offs.

[0003] To address the above problems, the present invention proposes a cross-temperature RC network based on non-volatile memory, which significantly improves the computing performance and reliability of the system by combining the dynamic characteristics at room temperature (300K) and the high stability at low temperature (77K). Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a low-power, high-performance, cross-temperature collaborative storage pool computing network based on temperature-controlled non-volatile access.

[0005] The present invention's temperature-controlled non-volatile access storage pool computing network adopts the following technical solutions:

[0006] The storage pool computing network includes a room temperature storage layer module, a temperature interface module and a low temperature reading layer module which are arranged in sequence;

[0007] Room-temperature storage layer module: Utilizes the dynamic characteristics (dynamic charge capture / release characteristics) of non-volatile memory (FeFET) to achieve short-term memory for fast signal processing, dynamic signal processing and short-term state evolution;

[0008] Temperature interface module: This module implements nonlinear mapping of signal characteristics between room temperature and low temperature environments, converting the dynamic output of the room temperature storage layer module into input that can be stably processed by the low temperature reading layer. This module enables smooth information connection across temperature domains and supports end-to-end gradient transfer.

[0009] Low-temperature readout layer module: Utilizes the stable characteristics (enhanced polarization stability) of non-volatile memory (FeFET) to achieve long-term memory for long-term data storage and weight update.

[0010] The room temperature storage layer module, temperature interface module and low temperature reading layer module are structurally connected in series, achieving decoupled coordination in terms of function.

[0011] The room temperature in the room temperature storage layer module is 300K. The low temperature in the low temperature reading layer module is 77K.

[0012] The low-temperature reading layer module performs weight update through the temperature-adaptive Hebbian learning rule.

[0013] The present invention's temperature-controlled non-volatile access reservoir computing network is used for image recognition, signal classification, speech processing, and other artificial intelligence computing scenarios, and has higher accuracy in complex tasks than traditional single-temperature systems.

[0014] In image recognition tasks, collaborative training is performed through forward propagation and back propagation. In the forward propagation stage, the room temperature reservoir layer processes dynamic signals, and the low temperature reading layer generates classification results. In the back propagation stage, the gradient is passed through the temperature interface module and used to update the network parameters, thereby achieving efficient training across temperature domains. In the forward propagation stage, the input image is extracted by CNN (convolutional neural network), and then dynamically processed by the room temperature reservoir layer, and then converted to the low temperature reading layer through the temperature interface module to generate classification results. In the back propagation stage, the gradient generated by the cross entropy loss is passed forward through the temperature interface module to update the weights and biases of modules such as the CNN and reservoir layer, while the low temperature reading layer uses the Hebbian learning rule to independently update its weights, thereby achieving efficient collaborative training across temperature domains.

[0015] The RC network of this invention is based on a non-volatile memory (FeFET) and adopts a multi-temperature zone structure. It achieves short-term dynamic response in the room temperature zone and long-term stable storage in the low temperature zone, thus taking into account both short-term memory (STM) and long-term memory (LTM) functions, effectively improving the accuracy, energy efficiency, and reliability of the neural network. By combining the dynamic characteristics at room temperature (300K) with the high stability at low temperature (77K), the system's computing performance and reliability are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1It is a structural block diagram of a storage pool computing network based on temperature-controlled non-volatile access of the present invention.

[0017] Figure 2 Schematic diagram of the gate stack structure of non-volatile memory (FeFET).

[0018] Figure 3 This is a comparison chart of the memory window (MW) of FeFET devices at different temperatures.

[0019] Figure 4 This is the retention performance test curve of non-volatile memory (FeFET) at 300K and 77K.

[0020] Figure 5 This is a graph showing the cycle endurance test of non-volatile memory (FeFET) at different temperatures.

[0021] Figure 6 Schematic diagram of the dynamic switching characteristics (STM) of non-volatile memory (FeFET) at 300K and 77K.

[0022] Figure 7 Schematic diagram of the weight retention characteristics (LTM) of non-volatile memory (FeFET) at 300K and 77K.

[0023] Figure 8 This is a performance comparison chart of the CIFAR-10 image recognition task based on the temperature-controlled non-volatile access storage pool computing network of the present invention.

[0024] In the figure: 1. substrate, 2. SiO2 layer, 3. HZO ferroelectric layer, 4. Al2O3 intermediate layer, 5. TiN top electrode. DETAILED DESCRIPTION

[0025] The present invention aims to design a storage pool computing network based on temperature-controlled non-volatile access, which achieves short-term dynamic response in the room temperature zone and long-term stable storage in the low temperature zone, thereby taking into account both short-term memory (STM) and long-term memory (LTM) functions, and effectively improving the accuracy, energy efficiency and reliability of the neural network.

[0026] like Figure 1 As shown, the storage pool computing network based on temperature-controlled non-volatile access of the present invention includes a room temperature storage layer module, a temperature interface module and a low temperature reading layer module which are arranged in sequence.

[0027] The room-temperature storage layer module is responsible for dynamic signal processing and short-term state evolution. The dynamic characteristics (dynamic charge capture / release characteristics) of non-volatile memory (FeFET) are utilized to realize short-term memory for fast signal processing.

[0028] The temperature interface module is used to realize the nonlinear mapping of signal characteristics between room temperature and low-temperature environments, connect the room-temperature storage layer and the low-temperature reading layer, convert the dynamic room-temperature signal into a stable low-temperature input feature, and realize effective information conversion and reverse gradient support across temperature domains.

[0029] The low-temperature readout layer module utilizes the enhanced polarization stability of non-volatile memory (FeFET) to achieve long-term memory for long-term data storage and weight update. Weight update is performed via a temperature-adaptive Hebbian learning rule.

[0030] The room-temperature reservoir module, temperature interface module, and low-temperature readout module are structurally connected in series. The room-temperature reservoir module is responsible for dynamic processing of input signals and simulating short-term plasticity. The temperature interface module acts as a bridge, implementing nonlinear mapping from room temperature to low-temperature states and supporting gradient transfer via backpropagation. The low-temperature readout layer provides stable classification output in low-temperature environments and performs local unsupervised weight updates using the Hebbian rule. Together, these three modules form a cross-temperature, functionally decoupled, and collaborative neural processing architecture.

[0031] The above-mentioned storage pool computing network is based on non-volatile memory (FeFET), and uses the temperature layered architecture to synergistically utilize the complementary characteristics of FeFET at different temperatures. The gate stack structure of the FeFET device is as follows: Figure 2 As shown, it includes a substrate 1, a SiO2 layer 2, a HZO ferroelectric layer 3, an Al2O3 intermediate layer 4 and a TiN top electrode 5, wherein the HZO ferroelectric layer is used to adjust the threshold voltage, and the Al2O3 intermediate layer is used to enhance the interface characteristics.

[0032] Depend on Figure 3 The comparison of the memory window (MW) of FeFET devices at different temperatures shows that in the low temperature characteristics (77K), the FeFET device exhibits a memory window (MW) of 8V, which is 45.5% higher than the 5.5V at room temperature (300K). Figure 4 The retention performance test curves of FeFET devices at 300K and 77K are given. It can be seen that in the low temperature characteristics (77K), the retention characteristics are excellent. 4 After 10 seconds, the storage window (MW) retention rate reaches 99.6%. Figure 5 The cycle durability test of FeFET devices at different temperatures shows that the durability is significantly enhanced in low temperature characteristics (77K). 7 After the first cycle, the degradation was only 0.4% (26.4% at room temperature).

[0033] FeFET devices have dynamic charge capture / release characteristics at room temperature (300K), enabling fast signal processing and short-term memory functions. Figure 6The dynamic switching characteristics (STM) of FeFET devices at 300K and 77K are given. Figure 7 The weight retention characteristics (LTM) of FeFET devices at 300K and 77K are given.

[0034] The temperature-controlled non-volatile access storage pool computing network of the present invention is suitable for image recognition, signal classification, speech processing and other artificial intelligence computing scenarios, and has a higher accuracy rate in complex tasks than traditional single-temperature systems. In image recognition tasks, training is achieved through forward propagation and backpropagation. In the forward propagation stage, after the input image is extracted by CNN (convolutional neural network), it is dynamically processed by the room temperature storage layer and then converted to the low temperature reading layer through the temperature interface module to generate the classification result. In the backpropagation stage, the cross entropy loss is calculated and the generated gradient is forwarded through the temperature interface module to update the weights and biases of modules such as CNN and the storage layer. At the same time, the low temperature reading layer performs Hebbian weight update.

[0035] Figure 8 The performance comparison of the cross-temperature RC network on the CIFAR-10 image recognition task is given. The training results on the CIFAR-10 dataset show that (see Figure 1 ), the hybrid temperature model of the present invention achieved a classification accuracy of 76.73%, significantly higher than the 41.65% classification accuracy of the room temperature model alone and the 23.69% classification accuracy of the low-temperature model alone. These results demonstrate that the present invention significantly improves classification accuracy while maintaining its low-temperature durability advantage.

Claims

1. A storage pool computing network based on temperature-controlled non-volatile access, characterized by: It includes a room temperature storage layer module, a temperature interface module and a low temperature reading layer module which are arranged in sequence; Room temperature storage layer module: Utilizes the dynamic characteristics of non-volatile memory to achieve short-term memory for fast signal processing, dynamic signal processing and short-term state evolution; Temperature interface module: This module implements nonlinear mapping of signal characteristics between room temperature and low temperature environments, converting the dynamic output of the room temperature storage layer module into input that can be stably processed by the low temperature reading layer. This module enables smooth information connection across temperature domains and supports end-to-end gradient transfer. Low-temperature readout layer module: Utilizes the stable characteristics of non-volatile memory to achieve long-term memory for long-term data storage and weight updates.

2. The temperature-controlled non-volatile access storage pool computing network according to claim 1, characterized in that: The room temperature in the room temperature storage layer module is 300K.

3. The temperature-controlled non-volatile access storage pool computing network according to claim 1, wherein: The low temperature of the low temperature reading layer module is 77K.

4. The temperature-controlled non-volatile access storage pool computing network according to claim 1, wherein: The low-temperature reading layer module performs weight update through the temperature-adaptive Hebbian learning rule.

5. The application of the temperature-controlled non-volatile access storage pool computing network according to any one of claims 1 to 4, characterized in that: Used for image recognition, signal classification, speech processing and other artificial intelligence computing scenarios.

6. The application of the temperature-controlled non-volatile access storage pool computing network according to claim 5, characterized in that Used in image recognition tasks, training is achieved through forward propagation and backpropagation. In the forward propagation stage, the room temperature storage layer processes dynamic signals, and the low temperature reading layer generates classification results; in the backpropagation stage, the parameters are updated through the temperature interface module.

7. The application of the temperature-controlled non-volatile access storage pool computing network according to claim 6, characterized in that: In the forward propagation stage, the input image is extracted with features by CNN, dynamically processed by the room temperature storage layer, and then converted to the low temperature reading layer through the temperature interface module to generate the classification result.

8. The application of the temperature-controlled non-volatile access storage pool computing network according to claim 6, characterized in that: In the back propagation phase, the cross entropy loss is calculated, and based on the generated gradient, the gradient is back propagated through the temperature interface module to update the weights and biases of modules such as the CNN and reserve layers, while the low-temperature read layer performs Hebbian weight updates.

Images