Deep photonic reservoir computing method based on two-layer micro-ring resonator coupling array

Abstract

The application belongs to the field of optical information processing and neuromorphic computing, and specifically discloses a deep photonic reservoir pool computing method based on a double-layer micro-ring resonator coupling array. The application constructs a 4*4*2 evanescent coupling double-layer micro-ring resonator photonic reservoir pool, which is a deep reservoir pool architecture with two layers of reservoir pool layers. Each layer is a 4*4 micro-ring coupling array, and each layer of micro-ring has different Q values, so as to design different fast and slow time scales to realize multi-time scale feature extraction of signal characteristics. The application optimizes the working state of the reservoir pool by adjusting and controlling the input power, global detuning and other working parameters. Only the output layer weight needs to be trained to complete the chaotic time series prediction, solving the problems of limited memory capacity of single-layer reservoir pool and insufficient prediction accuracy, effectively improving the feature representation ability and prediction accuracy of the reservoir pool, and having good on-chip integration and working stability.

Description

Technical Field

[0001] This invention relates to the fields of optical information processing and neuromorphic computing technology, specifically to a deep photon reservoir computing method based on a dual-layer micro-ring resonator coupled array. Background Technology

[0002] With the continuous growth in demand for artificial intelligence and high-speed information processing, the traditional von Neumann architecture, due to the separation of storage and computation, faces problems such as high data transfer overhead, high latency, and high energy consumption. Neuromorphic computing has attracted widespread attention due to its ability to simulate the efficient information processing mechanism of biological nervous systems. Reservoir computing, as an important branch of neuromorphic computing, has advantages such as simple training (requiring only training the output layer weights) and hardware-friendly implementation, showing great potential in tasks such as time series prediction and signal processing. Photonic reservoir computing utilizes the advantages of high bandwidth, high parallelism, and low power consumption of optical signals, making it suitable for high-speed time-domain information processing. Among various optical implementation platforms, silicon-based microring resonators have become an important candidate device for realizing photonic reservoirs due to their small device size, compatibility with CMOS processes, and rich nonlinear dynamic characteristics such as two-photon absorption and free carrier dispersion.

[0003] However, existing microring resonator reservoir schemes mostly employ single-layer structures or single-node time-delay multiplexing structures, which still have shortcomings in terms of memory capacity, dynamic representation dimensions, and multi-timescale feature extraction. For complex time series tasks, single-layer structures struggle to simultaneously accommodate both short-term transient information and long-term memory information. While existing research has proposed the concept of deep reservoirs, how to construct a deep photonic reservoir structure based on microring resonator arrays that possesses hierarchical feature extraction capabilities, can accommodate both fast and slow dynamic responses, and is suitable for on-chip integration remains a challenge. In microring resonator reservoirs, parameters such as input excitation intensity, global detuning between the input light and the resonant cavity, and resonant frequency deviations between devices significantly affect the reservoir's dynamic operating state and task performance. Without a proper design for the operating point, insufficient nonlinear mapping capability or memory degradation can easily occur, thus impacting the accuracy of time series prediction. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a deep photonic reservoir calculation method based on a dual-layer micro-ring resonator coupled array. This method solves the problems of limited structural memory capacity of existing micro-ring photonic reservoirs, difficulty in taking into account both short-term dynamic characteristics and long-term memory characteristics, insufficient prediction accuracy for complex time series, and difficulty in controlling the operating point of the reservoir.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for calculating a deep photon reservoir based on a dual-layer micro-ring resonator coupling array, comprising the following steps: S1. Generate Mackey-Glass chaotic time series: Calculate general parameters according to the reservoir to generate a continuous chaotic signal, and obtain a discrete time series after Euler numerical integration and downsampling, which serves as the original time series to be predicted. S2. Preprocessing the time series to be predicted: Normalize the discrete time series, then convert it into a continuous input excitation signal through a zero-order hold scheme, and adjust the input scaling factor to set the normalized input power in a range close to 1 so that the reservoir works in the effective nonlinear region. S3. Complete optical signal modulation and reservoir injection: A Mach-Zehnder modulator is used to load the continuous input excitation signal onto a 1550nm continuous wave laser carrier to obtain the modulated input optical signal, which is then injected into the reservoir of the double-layer micro-ring resonator through the input coupling waveguide. S4. Nonlinear mapping and state acquisition of the reservoir: Set the detuning matrix, coupling matrix and nonlinear response related parameters of the reservoir to enable the input optical signal to complete nonlinear dynamic mapping. Discretely sample the complex field amplitude of the reservoir at the corresponding time nodes, extract the real part and imaginary part of the complex field amplitude of each node, and construct a high-dimensional output feature matrix. S5. Training the linear readout layer: Using the output feature matrix as input and the Mackey-Glass discrete sequence with single-step lead as supervision signal, the output layer weight matrix is ​​solved based on the ridge regression method with the normalized mean square error as the loss function. After training, the matrix is ​​fixed. S6. Complete time series prediction and performance evaluation: Perform linear operation between the output feature matrix obtained after preprocessing the test set and the trained weight matrix to obtain the prediction output. Select memory capacity and normalized mean square error as evaluation indicators to complete the comprehensive evaluation of prediction performance.

[0006] Preferably, the double-layer microring resonator reservoir is a 4×4×2 evanescently coupled nonlinear microring resonator array structure, including an input layer, a double-layer reservoir layer, and an output layer; The dual-layer reservoir includes a first reservoir and a second reservoir, with different photon lifetime parameters set for the two layers. The linear decay rate of the first reservoir is much greater than that of the second reservoir. The first reservoir is used to extract fast dynamic features, while the second reservoir is used to extract slow dynamic features and provide a longer decay memory. The first and second reservoirs are connected by a bus waveguide to achieve unidirectional information transmission, allowing the input signal to propagate and evolve in the dual-layer reservoir in a unidirectional hierarchical manner.

[0007] Preferably, the column vector of the complex field amplitude composition of the microring resonators in the double-layer microring resonator reservoir is as follows: Its temporal evolution satisfies ; The detuning matrix is ​​represented as follows: , This represents the global detuning between the input laser and the average resonant frequency of the microring resonator array. This represents the random resonance deviation of the k-th microring resonator caused by manufacturing errors or structural inhomogeneities. The coupling matrix Represented in block matrix form , This represents the intra-layer coupling matrix within the first reserve layer. This represents the intra-layer coupling matrix within the second reserve layer. This is the interlayer transport matrix from the first reserve layer to the second reserve layer; The nonlinear response of the reservoir includes two-photon absorption and free carrier dispersion effects, and the nonlinear term of the k-th microring resonator is: ,in This represents the intensity-related loss term caused by two-photon absorption. The frequency shift term is caused by free carrier dispersion, and the free carrier density satisfies... , The carrier recombination rate, This is the carrier generation coefficient associated with two-photon absorption.

[0008] Preferably, in step S2, the continuous input excitation signal is obtained by amplitude scaling of the normalized discrete input sequence. The input scaling factor is used to adjust the input excitation intensity so that the continuous input excitation signal matches the nonlinear response range of the double-layer micro-ring resonator array. The normalized input power is the ratio of the input power to the reference nonlinear power, which corresponds to the reference operating point of the microring resonator array when it transitions from a weak nonlinear response to an effective nonlinear response. By adjusting the input scaling factor, the double-layer microring resonator reservoir can operate in different operating states in the linear region, the effective nonlinear region, or the overdrive region.

[0009] Preferably, the training of the linear readout layer is divided into two stages: state acquisition and output weight calculation. Specifically, the input time series is first input into the double-layer microring resonator reservoir, and the reserve states of each microring resonator node at each time step are acquired and sampled to construct an output feature matrix composed of the real and imaginary parts of the complex field amplitude of each node. Then, using the target output sequence as the supervision signal, the ridge regression method is used to solve for the output layer weights. The ridge regression training equation is as follows: ,in To output the feature matrix, The target set matrix, For ridge regularization factor, The identity matrix is ​​used; after training, the output layer weight matrix is ​​fixed and used for subsequent time series prediction.

[0010] Preferably, the operating parameters of the double-layer microring resonator reservoir include the input scaling factor, global detuning, and random resonance deviation between the microring resonators. The principle for adjusting the operating parameters is as follows: when the global detuning is within the preset optimal offset range (60-80 pm for Mackey-Glass prediction tasks), the amplitude of the random resonance deviation between the microring resonators is controlled to be less than 10 pm; when the global detuning is at an unfavorable operating point deviating from the aforementioned optimal offset range, a random resonance deviation with an amplitude of 50-100 pm is introduced, and the state space dimension of the reservoir can be enhanced by utilizing spectral diversity. By setting and adjusting the operating parameters, the double-layer microring resonator reservoir operates within a range that balances input coupling efficiency, intracavity field enhancement, nonlinear dynamic response, and attenuation memory capability, thereby improving the accuracy of time series prediction.

[0011] The beneficial effects of this invention are as follows: By constructing a dual-layer microring resonator array and setting different layer time scales, this invention enables the input signal to extract fast dynamic features in the first reservoir layer and slow dynamic features in the second reservoir layer, forming a longer decay memory, thereby improving the reservoir's ability to characterize complex time series; by setting a unidirectional coupling structure from the first reservoir layer to the second reservoir layer, hierarchical information transmission is achieved and inter-layer back interference is reduced; by introducing two-photon absorption effect and free carrier dispersion effect, the nonlinear mapping capability of the reservoir is enhanced; by adjusting the input scaling factor, global detuning amount, and random resonance deviation between microrings, the dual-layer deep photon reservoir operates at a more optimal dynamic operating point, improving the accuracy of time series prediction. Attached Figure Description

[0012] Figure 1 is a schematic diagram of the deep reservoir computing architecture based on microring resonators in an embodiment of the present invention; Figure 2 is a histogram of the frequency spacing between adjacent supermodes under different grid sizes in the embodiments of the present invention; Figure 3 shows the relationship between the total memory capacity of the double-layer micro-ring resonator reservoir and the size of each layer of the grid in an embodiment of the present invention. Figure 4 is a comparison diagram of the architecture of Mackey-Glass chaotic time series prediction in the embodiments of the present invention; Figure 5 shows the impact of normalized input power on the Mackey-Glass prediction task in this embodiment of the invention. Figure 6 shows the relationship between Mackey-Glass prediction performance and normalized input power under different nonlinear response models in the embodiments of the present invention. Figure 7 shows the impact of global detuning and random resonance deviation between micro-rings on the Mackey-Glass prediction task in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0014] This invention provides a technical solution: a method for calculating a deep photonic reservoir based on a dual-layer microring resonator coupled array. This invention constructs a dual-layer deep photonic reservoir based on a 4×4×2 evanescent coupled nonlinear microring resonator array to achieve Mackey-Glass chaotic time series prediction. All simulations are performed in Matlab based on temporal coupled-mode theory and the slowly varying envelope approximation. The method includes: preprocessing the time series to be processed to obtain a discrete input sequence; normalizing the discrete input sequence and converting it into a continuous input excitation signal using a zero-order hold method; loading the continuous input excitation signal onto a continuous optical laser carrier using a modulator to obtain a modulated input optical signal; injecting the input optical signal into a dual-layer deep photonic reservoir composed of a microring resonator array; performing nonlinear dynamic mapping and multi-timescale feature extraction on the input optical signal in the dual-layer deep photonic reservoir to obtain the reserve states of each node; sampling the reserve states and constructing an output feature matrix; training the output feature matrix using a linear readout layer to obtain output layer weights; and finally, outputting the target prediction result using the output layer weights.

[0015] The core steps of the simulation implementation of this invention are as follows: Step 1: Generate Mackey-Glass chaotic time series. Select the Mackey-Glass chaotic time series as the prediction target. Calculate general parameters according to the reservoir to generate a continuous chaotic signal. After Euler numerical integration and downsampling, a discrete time series is obtained, which serves as the original time series to be predicted.

[0016] Step 2: Preprocess the time series to be predicted. Normalize the discrete time series, then convert it into a continuous input excitation signal using a zero-order hold scheme. Adjust the input scaling factor to set the normalized input power close to 1, so that the reservoir operates in the effective nonlinear region.

[0017] Step 3: Complete optical signal modulation and reservoir injection. A Mach-Zehnder modulator is used to load the continuous input excitation signal onto a 1550nm continuous wave laser carrier to obtain the modulated input optical signal. This signal is then injected into the reservoir of a 4×4×2 double-layer microring resonator through an input coupling waveguide. The lifetimes of the two layers are set according to the requirements of fast and slow dynamic feature extraction, and unidirectional information transmission between the layers is achieved through a dedicated bus waveguide.

[0018] Step 4: Implement nonlinear mapping and state acquisition of the reservoir. Set the detuning matrix, coupling matrix, and nonlinear response parameters of the reservoir to enable the input optical signal to complete nonlinear dynamic mapping. Discretely sample the complex field amplitude of the reservoir at corresponding time nodes, extract the real and imaginary parts of the complex field amplitude at each node, and construct a high-dimensional output feature matrix.

[0019] Step 5: Train the linear readout layer. Using the output feature matrix as input and the Mackey-Glass discrete sequence with single-step lead as the supervision signal, the output layer weight matrix is ​​solved based on the ridge regression method with the normalized mean square error as the loss function. After training, this matrix is ​​fixed.

[0020] Step 6: Complete time series prediction and performance evaluation. Perform a linear operation between the output feature matrix obtained after preprocessing the test set and the trained weight matrix to obtain the predicted output. Select memory capacity and normalized mean square error as evaluation metrics to complete a comprehensive evaluation of the prediction performance.

[0021] Furthermore, the double-layer deep photon reservoir consists of multiple evanescently coupled nonlinear microring resonators, such as... Figure 1 As shown, it includes an input layer, a dual-layer storage layer, and an output layer. The dual-layer storage layer comprises a first storage layer and a second storage layer. The first storage layer is used to extract fast dynamic features from the input signal, and the second storage layer is used to extract slow dynamic features from the input signal and provide longer decay memory. Figure 2 As shown, this is a histogram of the frequency spacing between adjacent supermodes under different grid sizes per layer, reflecting the impact of reservoir size on spectral diversity. Figure 3 The figure shows the relationship between the total memory capacity of the reservoir of the double-layer microring resonator and the grid size of each layer, illustrating the correlation between reservoir size and memory capacity. Unidirectional information transmission is achieved between the first and second reservoir layers via an interlayer coupling waveguide, allowing the input signal to be transmitted to the second reservoir layer for deeper dynamic evolution after initial dynamic mapping in the first reservoir layer.

[0022] Furthermore, the column vector comprising the complex field amplitudes of all microring resonators in the dual-layer deep photon reservoir is: Its temporal evolution satisfies ,in, This is the detuning matrix, used to represent the global detuning between the input laser and the average resonant frequency of the microring resonator array, as well as the random resonance deviation between each microring resonator. It is a linear decay rate matrix. This is the coupling matrix, used to describe the intra-layer and inter-layer coupling between microring resonators. It is a nonlinear response term. For the input coupling vector, This is the input excitation signal.

[0023] The detuning matrix is ​​represented as follows: ,in, This represents the global detuning between the input laser and the average resonant frequency of the microring resonator array. Indicates the first Random resonance deviations in a microring resonator due to manufacturing errors or structural inhomogeneities. The coupling matrix... Represented in block matrix form as ,in, This represents the intra-layer coupling matrix within the first reserve layer. This represents the intra-layer coupling matrix within the second reserve layer. This is the interlayer transfer matrix from the first reservoir layer to the second reservoir layer. Through this block matrix structure, the input signal propagates and evolves in a unidirectional hierarchical manner within the two reservoir layers.

[0024] The dual-layer deep photon reservoir employs a 4×4×2 evanescently coupled nonlinear microring resonator array structure. Both reservoir layers are 4×4 evanescently coupled microring resonator arrays, achieving unidirectional information transmission through interlayer coupling waveguides. The two reservoir layers are configured with different quality factor parameters: the first reservoir layer uses a low quality factor design, corresponding to a fast photon attenuation response, while the second reservoir layer uses a high quality factor design, corresponding to a slow photon attenuation response. Specifically, the low quality factor first reservoir layer is used to capture high-frequency details and fast transient nonlinear features in the input signal, while the high quality factor second reservoir layer is used to provide longer attenuation memory and extract long-term dynamic features from the input signal. By designing differentiated Q-values ​​for the two reservoir layers, a fast-slow co-processing multi-timescale processing mechanism is constructed to achieve joint processing of the input time series across multiple timescales.

[0025] Furthermore, the nonlinear response of the bilayer deep photon reservoir includes two-photon absorption and free carrier dispersion effects. For the first... A microring resonator, whose nonlinear term is expressed as: ,in, This represents the intensity-related loss term caused by two-photon absorption. This represents the frequency shift term caused by free carrier dispersion. The free carrier density satisfies... ,in, Indicates the carrier recombination rate. This represents the carrier generation coefficient associated with two-photon absorption. Through the combined effect of two-photon absorption and free carrier dispersion, the microring resonator array possesses the ability to perform high-dimensional nonlinear mapping of the input signal.

[0026] The input excitation signal is a normalized discrete input sequence that has been amplitude-scaled and then input into the two-layer deep photon reservoir. Its expression is: ,in, This represents the normalized discrete input sequence. Indicates the input scaling factor. The input scaling factor indicates the sampling and holding time. It is used to adjust the input excitation intensity so that the continuous input excitation signal matches the nonlinear response range of the dual-layer microring resonator array. To characterize the relative relationship between the input excitation intensity and the nonlinear operating threshold of the microring resonator, a normalized input power is introduced, which is the ratio of the input power to the reference nonlinear power. The reference nonlinear power corresponds to the reference operating point where the microring resonator array transitions from a weak nonlinear response to an effective nonlinear response. By setting the normalized input power, the dual-layer deep photon reservoir can operate in different states within the linear region, the effective nonlinear region, or the overdriven region.

[0027] Furthermore, the training of the dual-layer deep photon reservoir is divided into two stages: state acquisition and output weight calculation. First, the input time series is input into the dual-layer deep photon reservoir, and the reserve states of each microring resonator node at each time step are acquired. The reserve states are then sampled to construct the output feature matrix. For the time series prediction task, the output feature matrix is ​​composed of the real and imaginary parts of the complex amplitude field of each node. Then, using the target output sequence as the supervision signal, the ridge regression method is used to solve for the output layer weights, thereby completing the training of the linear readout layer. After training, the trained output layer weights are fixed and used for subsequent time series prediction tasks.

[0028] The operating parameters of the dual-layer deep photon reservoir include the input scaling factor, global detuning, and random resonance deviation between microring resonators. By setting and adjusting these operating parameters, the dual-layer deep photon reservoir can operate within a range that balances input coupling efficiency, intracavity field enhancement, nonlinear dynamic response, and attenuation memory capability, thereby improving the accuracy of time series prediction.

[0029] The time series prediction task is a chaotic time series prediction task. After inputting the time series to be predicted into a two-layer deep photon reservoir, the predicted value of the future step size of the time series is obtained by using the output layer.

[0030] Numerical simulation To verify the effectiveness of the proposed deep photon reservoir calculation method based on a double-layer microring resonator array in time series prediction tasks, three sets of comparative simulation experiments were set up, such as... Figure 4 The figures show a single-layer 4×4 grid microring resonator reservoir, a double-layer 4×4×2 grid microring resonator reservoir with equal photon lifetimes, and the double-layer 4×4×2 grid fast-slow photon lifetime microring resonator reservoir proposed in this invention. All other process parameters and simulation parameters are kept consistent across the three sets of experiments; only the reservoir hierarchy and photon lifetime settings are differentiated.

[0031] Simulation results show that, in terms of memory capacity, the proposed dual-layer 4×4×2 grid fast-slow photon lifetime microring resonator reservoir has a higher total memory capacity. In grid size comparisons, the 4×4 grid achieves a better balance between prediction performance, memory capacity, and hardware implementation complexity; when the grid is further expanded to 5×5, the system size increases but the performance improvement is limited, and the chip area and control complexity increase significantly.

[0032] In the Mackey-Glass chaotic time series prediction task, the single-layer reservoir exhibits the highest normalized mean square error (MSE). The MSE of the double-layer equal-photon lifetime reservoir is significantly lower than that of the single-layer structure. The proposed double-layer fast-slow photon lifetime reservoir further reduces the MSE to 0.0039. Furthermore, after optimizing the global detuning parameters, the MSE can be further reduced to the order of 1.8 × 10^(-3). These results demonstrate that this invention does not merely improve prediction performance by increasing the number of nodes, but rather enhances the reservoir's ability to characterize the dynamic features of chaotic time series through a double-layer hierarchical structure and joint processing of fast and slow time scales.

[0033] Figure 5 shows the impact of normalized input power on the Mackey-Glass prediction task, illustrating the optimal input power operating point of the reservoir; Figure 6The figure shows the relationship between the Mackey-Glass prediction performance and the normalized input power under different nonlinear response models, comparing the prediction effects of various nonlinear mechanisms. In the simulations of the nonlinear response mechanism and the input power operating point, the system exhibits superior prediction performance in the effective nonlinear region where the normalized input power is close to 1 when two-photon absorption and free carrier dispersion work together. In contrast, the prediction performance under the single two-photon absorption mechanism and the Kerr nonlinear mechanism is weaker. Among them, the Kerr nonlinear mechanism is more prone to overdrive when the normalized input power is large, leading to a significant increase in prediction error. The above results indicate that the nonlinear mechanism of two-photon absorption and free carrier dispersion used in this invention is beneficial for forming dynamic mapping characteristics more suitable for time series prediction.

[0034] In operating parameter tuning and robustness simulation, when the global detuning is at an optimal value, the random resonance deviation between microring resonators should be kept within a small range, at which point the system can achieve a low prediction error. When the global detuning is at an unfavorable operating point, appropriately introducing random resonance deviation can improve the effective excitation conditions of some nodes, thereby reducing the prediction error. For example... Figure 7 The figure shows the impact of global detuning and random resonance deviation between microrings on the Mackey-Glass prediction task, illustrating the regulation of key operating parameters of the reservoir. Further simulations show that, under certain conditions of global detuning fluctuations and random resonance deviations, the proposed double-layer microring resonator deep photon reservoir can still maintain a low prediction error, indicating that this technical solution has a certain adaptability and robustness to silicon-based chip manufacturing errors and operating point shifts.

[0035] In summary, this invention effectively enhances the photon reservoir's ability to process complex time-series tasks by constructing a dual-layer micro-ring resonator array, setting fast and slow time scales, introducing a joint nonlinear mechanism of two-photon absorption and free carrier dispersion, and combining the control of operating parameters such as normalized input power, global detuning, and random resonance bias. It has good prediction accuracy and promising prospects for engineering implementation.

[0036] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0037] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0038] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0039] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0040] The terms "first" and "second" used in the embodiments are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0041] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for calculating deep photon reservoirs based on a dual-layer micro-ring resonator coupled array, characterized in that, Includes the following steps: S1. Generate Mackey-Glass chaotic time series: Calculate general parameters according to the reservoir to generate a continuous chaotic signal, and obtain a discrete time series after Euler numerical integration and downsampling, which serves as the original time series to be predicted. S2. Preprocessing the time series to be predicted: Normalize the discrete time series, then convert it into a continuous input excitation signal through a zero-order hold scheme, and adjust the input scaling factor to set the normalized input power in a range close to 1 so that the reservoir works in the effective nonlinear region. S3. Complete optical signal modulation and reservoir injection: A Mach-Zehnder modulator is used to load the continuous input excitation signal onto a 1550nm continuous wave laser carrier to obtain the modulated input optical signal, which is then injected into the reservoir of the double-layer micro-ring resonator through the input coupling waveguide. S4. Nonlinear mapping and state acquisition of the reservoir: Set the detuning matrix, coupling matrix and nonlinear response related parameters of the reservoir to enable the input optical signal to complete nonlinear dynamic mapping. Discretely sample the complex field amplitude of the reservoir at the corresponding time nodes, extract the real part and imaginary part of the complex field amplitude of each node, and construct a high-dimensional output feature matrix. S5. Training the linear readout layer: Using the output feature matrix as input and the Mackey-Glass discrete sequence with single-step lead as supervision signal, the output layer weight matrix is ​​solved based on the ridge regression method with the normalized mean square error as the loss function. After training, the matrix is ​​fixed. S6. Complete time series prediction and performance evaluation: Perform linear operation between the output feature matrix obtained after preprocessing the test set and the trained weight matrix to obtain the prediction output. Select memory capacity and normalized mean square error as evaluation indicators to complete the comprehensive evaluation of prediction performance.

2. The method for calculating a deep photon reservoir based on a dual-layer micro-ring resonator coupled array according to claim 1, characterized in that: The dual-layer microring resonator reservoir is a 4×4×2 evanescently coupled nonlinear microring resonator array structure, including an input layer, a dual-layer reservoir layer, and an output layer. The dual-layer reservoir includes a first reservoir and a second reservoir, with different photon lifetime parameters set for the two layers. The linear decay rate of the first reservoir is much greater than that of the second reservoir. The first reservoir is used to extract fast dynamic features, while the second reservoir is used to extract slow dynamic features and provide a longer decay memory. The first and second reservoirs are connected by a bus waveguide to achieve unidirectional information transmission, allowing the input signal to propagate and evolve in the dual-layer reservoir in a unidirectional hierarchical manner.

3. The method for calculating deep photon reservoirs based on a dual-layer micro-ring resonator coupling array according to claim 1, characterized in that: The column vector of the complex field amplitude composition of the microring resonators in the double-layer microring resonator reservoir is as follows: Its temporal evolution satisfies ; The detuning matrix is ​​represented as follows: , This represents the global detuning between the input laser and the average resonant frequency of the microring resonator array. This represents the random resonance deviation of the k-th microring resonator caused by manufacturing errors or structural inhomogeneities. The coupling matrix Represented in block matrix form , This represents the intra-layer coupling matrix within the first reserve layer. This represents the intra-layer coupling matrix within the second reserve layer. This is the interlayer transport matrix from the first reserve layer to the second reserve layer; The nonlinear response of the reservoir includes two-photon absorption and free carrier dispersion effects, and the nonlinear term of the k-th microring resonator is: ,in This represents the intensity-related loss term caused by two-photon absorption. The frequency shift term is caused by free carrier dispersion, and the free carrier density satisfies... , The carrier recombination rate, This is the carrier generation coefficient associated with two-photon absorption.

4. The method for calculating a deep photon reservoir based on a dual-layer micro-ring resonator coupling array according to claim 1, characterized in that: In step S2, the continuous input excitation signal is obtained by amplitude scaling of the normalized discrete input sequence. The input scaling factor is used to adjust the input excitation intensity so that the continuous input excitation signal matches the nonlinear response range of the double-layer micro-ring resonator array. The normalized input power is the ratio of the input power to the reference nonlinear power, which corresponds to the reference operating point of the microring resonator array when it transitions from a weak nonlinear response to an effective nonlinear response. By adjusting the input scaling factor, the double-layer microring resonator reservoir can operate in different operating states in the linear region, the effective nonlinear region, or the overdrive region.

5. The method for calculating a deep photon reservoir based on a dual-layer micro-ring resonator coupling array according to claim 1, characterized in that: The training of the linear readout layer is divided into two stages: state acquisition and output weight calculation. Specifically, the input time series is first input into the double-layer microring resonator reservoir, the reserve state of each microring resonator node at each time is acquired and sampled, and the output feature matrix is ​​constructed by the real part and imaginary part of the complex field amplitude of each node. Then, using the target output sequence as the supervision signal, the ridge regression method is used to solve for the output layer weights. The ridge regression training equation is: ,in To output the feature matrix, The target set matrix, For ridge regularization factor, It is the identity matrix; After training, the output layer weight matrix is ​​fixed and used for subsequent time series prediction.

6. The method for calculating a deep photon reservoir based on a dual-layer micro-ring resonator coupling array according to claim 1, characterized in that: The operating parameters of the double-layer microring resonator reservoir include the input scaling factor, global detuning, and random resonance deviation between the microring resonators. The principle for adjusting the operating parameters is as follows: when the global detuning is within the preset optimal offset range, the amplitude of the random resonance deviation between the microring resonators is controlled to be less than 10 pm; when the global detuning is at an unfavorable operating point deviating from the aforementioned optimal offset range, a random resonance deviation with an amplitude of 50-100 pm is introduced, and the state space dimension of the reservoir can be enhanced by utilizing spectral diversity. By setting and adjusting the operating parameters, the double-layer microring resonator reservoir can operate within a range that balances input coupling efficiency, intracavity field enhancement, nonlinear dynamic response, and attenuation memory capability, thereby improving the accuracy of time series prediction.

Images