Quantum ring surface code decoder and method

Abstract

The invention discloses a quantum toroidal code decoder and a quantum toroidal code decoding method. A self-attention U-net quantum error correction decoder obtained through training is used for decoding a toroidal code. The SU-NetQD is composed of two parts, namely a low-level decoder and a high-level decoder. Wherein the low-level decoder uses the extracted error syndrome as input and predicts a recovery chain operation, and the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts a logic error introduced by the low-level decoder. When the decoder successfully decodes and only when the recovery chain predicted by the low-level decoder can completely disappear the error syndrome, and the logic error predicted by the high-level decoder conforms to the combined action effect of the recovery chain and the original error. According to the method, the denoising function can be efficiently and quickly implemented for the toroidal code under complex noise, the original information of the quantum bit is prevented from being interfered, and the performance is superior to that of a traditional MWPM decoder under various conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of quantum computing, and in particular relates to a quantum torus code decoder and method. Background Art

[0002] Quantum computing has become one of the most hotly debated topics in the current scientific and technological community due to its theoretical super computing power. For certain complex problems, researchers have proposed many efficient quantum algorithms that can provide quadratic or even exponential acceleration compared to classical computing. However, due to the existence of quantum noise, existing quantum computers still exhibit high error rates, which limits the scale and complexity of feasible quantum algorithms. Therefore, quantum error correction, which aims to detect and correct quantum bit errors in quantum systems, remains an extremely important and active research field. Quantum error correction is a necessary step to achieve reliable large-scale quantum computing.

[0003] Surface codes are one of the most popular and hardware-friendly error correction codes. Traditional surface code decoding methods, such as minimum weight perfect matching (MWPM) and union find (UF), face problems with decoding accuracy and scalability. Therefore, designing a new low-latency, high-precision, and scalable decoder has become a scientific problem that needs to be solved urgently. Recently, machine learning technology based on neural networks has attracted widespread attention due to some good properties. However, existing solutions, such as MLP, CNN, RNN, etc., do not consider the structural characteristics of surface codes, and their decoding accuracy and scalability need to be improved. At the same time, some complex network architectures, such as Transformer, still have high requirements for the size of the training data set, which is difficult to meet for scarce quantum error correction data.

[0004] In summary, the problems and defects of the existing technical solutions can be summarized as follows:

[0005] 1. Scale limitation: The minimum weight perfect matching algorithm is the most popular traditional surface code decoding method. However, it is limited by the time complexity of the algorithm and is unable to perform error correction tasks in large-scale quantum systems.

[0006] 2. Accuracy limitation: Although the minimum weight perfect matching algorithm has very high error correction capabilities at low error rates, it performs poorly in complex correlated noise and high error rate scenarios, which can easily lead to logical errors and low decoding accuracy.

[0007] 3. High computational overhead: Many neural network decoder-based models, such as multi-layer perceptrons, have fixed input and output sizes. Therefore, changes in the surface code distance will require retraining a completely new model, resulting in considerable overhead.

[0008] 4. High data requirements: Neural network models based on the Transformer structure often require more training data to achieve better decoding effects due to the lack of certain translation invariance and local perception.

[0009] 5. Model structure design defects: When the existing neural network decoder model unfolds the surface code into a two-dimensional plane graphic, it does not take into account the torus structure characteristics of the surface code itself and the relationship between the error chain and the original lattice and the dual lattice, which leads to low accuracy in predicting edge errors.

[0010] 6. Poor decoder scalability: Most existing decoders focus only on low-level decoders that detect the exact location of errors in data qubits, while ignoring the role of high-level decoders that correct logical errors. A systematic and scalable decoder has not yet been proposed. Summary of the invention

[0011] In order to overcome the shortcomings of the prior art, the present invention provides a quantum torus code decoder and method, which uses a trained self-attention U-net quantum error correction decoder (SU-NetQD) to decode the torus code. SU-NetQD consists of two parts: a low-level decoder and a high-level decoder. Among them, the low-level decoder uses the extracted error syndrome as input and predicts the recovery chain operation, and the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical error introduced by the low-level decoder. The decoder decodes successfully if and only if the recovery chain predicted by the low-level decoder can completely eliminate the error syndrome, and the logical error predicted by the high-level decoder conforms to the joint effect of the recovery chain and the original error. The present invention can efficiently and quickly implement the denoising function for torus codes under complex noise, keep the original information of the quantum bit undisturbed, and outperform the traditional MWPM decoder in various situations.

[0012] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0013] A quantum torus code decoder, comprising a low-level decoder and a high-level decoder;

[0014] The low-level decoder adopts a self-attention U-net (SU-Net) model, which consists of a two-dimensional U-net structure for extracting information within the lattice, a three-dimensional self-attention block for extracting the dependency of different types of error syndromes between lattices, and a residual block;

[0015] The advanced decoder reuses three SU-Net models. First, two SU-Net models are used to extract the original error syndrome and the context information of the recovery chain respectively. Then, the intermediate features are added to fuse the information of the two and input into another SU-Net model for final feature extraction. Finally, a global average pooling layer is used to average all elements in each channel and directly used for classification.

[0016] The low-level decoder uses the error syndrome as input and predicts the recovery chain operation; the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical errors introduced by the low-level decoder.

[0017] A quantum torus code decoding method comprises the following steps:

[0018] Step 1: Use the MWPM algorithm to decode the error syndrome generated by the collected torus code and the recovery operation that needs to be performed in a noisy environment, and correct the recovery operation predicted by MWPM by detecting the logical quantum state of the torus code;

[0019] Step 2: Build low-level decoder and high-level decoder;

[0020] The SU-Net used in low-level decoding consists of a 2D U-Net structure that extracts information within the lattice, and a 3D self-attention block and residual block that extracts dependencies between lattices for different types of error syndromes;

[0021] The advanced decoder reuses three SU-Net models. First, two SU-Nets are used to extract the original error syndrome and the context information of the recovery chain respectively. Then, the intermediate features are added to fuse the information of the two and input into another SU-Net for the final feature extraction. Finally, a global average pooling layer is used to average all elements in each channel and directly used for classification.

[0022] Step 3: Use the error syndrome in step 1 as the training dataset feature vector and the recovery chain operation as the dataset label, train the low-level decoder built in step 2 to make it converge, and record the recovery chain operation predicted by the low-level decoder for the last time, as well as the introduced logical errors;

[0023] Step 4: Take the error syndrome together with the recovery chain operation predicted by the low-level decoder in step 3 as the feature vector of the training dataset, take the logical errors occurring on the torus code as the dataset label, and train the model of the high-level decoder constructed in step 2 to make it converge;

[0024] Step 5: For new errors generated on the torus code under the same noise, the low-level decoder uses the extracted error syndrome as input and iteratively predicts the recovery chain operation, and the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical errors introduced by the low-level decoder;

[0025] The final decoding is successful if and only if the recovery chain predicted by the low-level decoder can make the error syndrome completely disappear, and the logical error predicted by the high-level decoder is consistent with the joint effect of the recovery chain and the original error.

[0026] Furthermore, the recovery chain predicted by the low-level decoder once cannot completely eliminate the original error syndrome; an iterative decoding strategy is used to first use the currently predicted recovery chain to eliminate part of the error syndrome, and then the remaining error syndrome is input into the low-level decoder again for prediction, and this operation is repeated until the error syndrome disappears or the pre-set maximum number of predictions is reached.

[0027] A computer program, wherein the computer program enables a computer to execute the above-mentioned quantum torus code decoding method.

[0028] An electronic device comprises: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the above-mentioned quantum torus code decoding method.

[0029] A computer-readable storage medium stores a computer program, which implements the above-mentioned quantum torus code decoding method when executed by a processor.

[0030] A chip includes: a processor, which is used to call and run a computer program from a memory, so that a device equipped with the chip executes the above-mentioned quantum torus code decoding method.

[0031] A computer program product, comprising a computer storage medium, wherein the computer storage medium stores a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the above-mentioned quantum torus code decoding method is implemented.

[0032] The beneficial effects of the present invention are as follows:

[0033] First, the number of downsampling and upsampling layers in the SU-Net network structure of the present invention can be automatically adjusted according to the specific accuracy requirements and computational complexity. The decoding accuracy of SU-NetQD under different physical error rates and code distances is much higher than that of the MWPM decoder, and the average decoding time is significantly reduced, regardless of the torus code size. The measured threshold value is increased from 0.095 to 0.231. At the same time, the advanced decoder of the present invention can also significantly improve the performance of the MWPM decoder.

[0034] Second, the present invention can implement the denoising function efficiently and quickly for torus codes under complex noise, keeping the original information of the quantum bits undisturbed, and outperforming the traditional MWPM decoder in various situations.

[0035] Third, the present invention is universal and practical, and is applicable to a variety of quantum hardware platforms in the NIST era, including superconducting quantum, ion trap, optical quantum, topological quantum, etc. The present invention is an important aspect of quantum information processing---quantum error correction, which can be integrated into quantum computing software packages to improve the decoding accuracy of torus codes and enhance their practicality.

[0036] Fourth, the technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0037] The existing torus code decoding scheme has certain defects in decoding accuracy, decoding time and scalability. The torus code decoding scheme proposed in the present invention has the advantages of high accuracy, portability, and easy scalability, and is suitable for decoding of recent quantum torus codes. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of a torus code structure and logical operations thereon provided by an embodiment of the present invention;

[0039] Figure 2 is an error syndrome measurement circuit diagram provided by an embodiment of the present invention;

[0040] Figure 3 is a schematic diagram of quantum error correction provided by an embodiment of the present invention;

[0041] Figure 4 It is a schematic diagram of the overall decoding process of SU-NetQD provided by an embodiment of the present invention;

[0042] Figure 5 is a schematic diagram of an iterative decoding process of a low-level decoder provided by an embodiment of the present invention;

[0043] Figure 6 It is a schematic diagram of the mapping relationship between quantum bits on a torus code and classical data provided by an embodiment of the present invention;

[0044] Figure 7is a schematic diagram of the structure of a low-level decoder provided by an embodiment of the present invention;

[0045] Figure 8 is a schematic diagram of the structure of an advanced decoder provided by an embodiment of the present invention;

[0046] Fig. 9 It is a schematic diagram of the decoding performance comparison results of SU-NetQD and MWPM under different noises and error rates provided by an embodiment of the present invention;

[0047] Fig.10 is the threshold error rate p provided in the embodiment of the present invention c Schematic diagram of the relationship with deviation η. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] The purpose of the present invention is to provide a high-precision, scalable, and compatible efficient quantum torus code decoder design method, which is based on the U-Net structure and multi-head self-attention mechanism, adopts the idea of ​​hierarchical design and gradual refinement, and effectively solves the problems existing in the prior art.

[0050] The technical solution adopted by the present invention is to use the trained SU-NetQD to decode the torus code. SU-NetQD consists of two parts: a low-level decoder and a high-level decoder. Among them, the low-level decoder uses the extracted error syndrome as input and predicts the recovery chain operation, and the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical error introduced by the low-level decoder. The decoder decodes successfully if and only if the recovery chain predicted by the low-level decoder can completely eliminate the error syndrome, and the logical error predicted by the high-level decoder conforms to the joint effect of the recovery chain and the original error.

[0051] Rearranging the data according to the position of the lattice where the quantum bit is located helps to extract potential features. At the same time, in order to make full use of the torus structure characteristics of the torus code, the traditional zero filling scheme is abandoned and the torus filling method is used instead, which can effectively improve the ability to correct edge errors.

[0052] The low-level decoder adopts an iterative decoding strategy, predicts in batches, and gradually refines the decoding results. In each iteration, the low-level decoder predicts the recovery chain based on the syndrome remaining after the previous round of decoding, and further eliminates the syndrome. At the same time, according to the number of error syndromes before and after the iteration, the low-level decoder trained under different error rates can be used for targeted decoding.

[0053] The low-level decoder adopts a novel SU-Net structure, which not only effectively utilizes the inductive bias properties of convolution, but also captures the long-range dependencies in the surface code, achieving a good balance between computational complexity and decoding accuracy, and requiring less training data.

[0054] The advanced decoder reuses three SU-Net structures, analogy multi-label classification problems, advanced decoder input syndrome and low-level decoder prediction recovery chain, prediction logic errors. The advanced decoder can be very conveniently combined with the traditional decoder and greatly improve the decoding accuracy of the traditional decoder.

[0055] SU-NetQD adopts a transfer learning strategy and can perform transfer learning on toroidal codes of different code distances. It not only avoids the overhead caused by repeated training, but also significantly enhances the performance of the model on toroidal codes with small code distances.

[0056] The torus code decoding solution provided by the embodiment of the present invention generally includes the following steps:

[0057] 1. The designed low-level decoder and high-level decoder networks are trained by using the existing torus code error correction data under specific noise and error rate, so that the low-level decoder can accept the input of error syndrome and output the predicted recovery chain operation, and the high-level decoder can simultaneously accept the recovery chain predicted by the low-level decoder and the original error syndrome and output the predicted logical error.

[0058] 2. For new errors generated on the toroidal code in a similar noise environment, first use the low-level decoder to iteratively decode the errors to obtain the initial recovery chain operation, then use the high-level decoder to predict the logical errors that may be introduced during the decoding process, use the corresponding logical operations to correct them, and finally obtain the recovery chain that needs to be acted on the toroidal code.

[0059] like Figure 1 : This is a schematic diagram of the torus code structure and the logical operations thereon provided by an embodiment of the present invention; the torus code is a square lattice with periodic boundary conditions, and each side of the torus has a quantum bit (represented by a circle). The solid line represents the original lattice, and the dotted line represents the dual lattice; the operations on the torus code are divided into trivial rings and non-trivial rings, where the trivial ring is usually local and can be reduced to a single vertex through local operations, and has no effect on the system. The non-trivial ring surrounds the entire lattice and cannot be reduced to a point through continuous deformation, which will cause logical changes in the system.

[0060] like Figure 3It is a schematic diagram of quantum error correction provided by an embodiment of the present invention; the left figure is an X-type error occurring on a torus code under bit flip noise; the middle sub-figure represents a correct recovery operation because it forms a trivial ring with the error chain; the right figure represents an incorrect recovery operation because it forms a non-trivial ring with the error chain, which will cause a logical error in the system.

[0061] Specifically include:

[0062] Step 1: Collect the error syndromes generated by the torus code and the recovery operations that need to be performed in a specific noise environment. The MWPM algorithm can be used to decode and collect them, and the recovery operations predicted by the MWPM can be corrected by detecting the logical quantum state of the torus code. Figure 2 Shown is the error syndrome measurement circuit diagram; the left figure is a Z-type error detection circuit; the right figure is an X-type error detection circuit.

[0063] Step 2: Figure 7 , Figure 8 The model structure shown builds a low-level decoder and a high-level decoder. The SU-Net used in the low-level decoding consists of a two-dimensional U-Net structure that extracts information within the lattice, and a three-dimensional self-attention block and residual block that extracts the dependencies of different types of error syndromes between lattices. The high-level decoder reuses three SU-Net models. It first uses two SU-Nets to extract the original error syndrome and the context information of the recovery chain respectively, then adds the obtained intermediate features to fuse the information of the two, and inputs them into another SU-Net for the final feature extraction. Finally, the global average pooling layer is used to average all elements in each channel and directly used for classification.

[0064] Step 3: Use the error syndrome in step 1 as the feature vector of the training dataset and the recovery chain operation as the dataset label to train the low-level decoder built in step 2 to make it converge. Record the recovery chain operation predicted by the low-level decoder for the last time and the introduced logical errors.

[0065] In step 4, the error syndrome and the recovery chain operation predicted by the low-level decoder in step 3 are used as the feature vector of the training data set, and the logical errors occurring on the torus code are used as the data set labels. The high-level decoder built in step 2 is trained to converge.

[0066] Step 5: Figure 4As shown in the figure, for new errors generated on the torus code under the same noise, the low-level decoder uses the extracted error syndrome as input and iteratively predicts the recovery chain operation, while the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical error introduced by the low-level decoder. The decoder decodes successfully if and only if the recovery chain predicted by the low-level decoder can completely eliminate the error syndrome, and the logical error predicted by the high-level decoder conforms to the joint effect of the recovery chain and the original error.

[0067] Due to the complexity of the errors, the recovery chain predicted by the low-level decoder once may not completely eliminate the original error syndrome. In this case, an iterative decoding strategy can be used to first use the currently predicted recovery chain to eliminate part of the error syndrome, and then input the remaining error syndrome into the low-level decoder again for prediction, and repeat this operation until the error syndrome disappears or the maximum number of predictions set in advance is reached. In view of this feature, the present invention recommends training multiple low-level decoders to learn the potential noise distribution under different error rates. For example, after the low-level decoder eliminates part of the error syndrome, the remaining error distribution is similar to the error distribution under low noise, so the decoder trained under low noise can be used to perform more accurate decoding. If the error syndrome disappears completely, the final recovery chain operation is obtained by XORing the recovery chain predicted at each stage. Figure 5 It is a schematic diagram of the iterative decoding process of the low-level decoder provided by an embodiment of the present invention; the low-level decoder receives the original error syndrome as input, and each iteration predicts the recovery chain based on the error syndrome remaining after the previous round of decoding, and further eliminates the error syndrome; the final output of the low-level decoder is the XOR of the recovery chain predicted at each stage.

[0068] Even under the same noise model, toroidal codes of different code distances will show different anti-interference capabilities. In order to decode toroidal codes of new code distances, traditional methods usually require retraining at a huge computational cost. SU-NetQD adopts a transfer learning strategy, which uses the similarity of potential error distribution between toroidal codes of different code distances under the same physical error rate to fine-tune the pre-trained model on the surface code of the new distance, which can effectively shorten the training time. At the same time, using transfer learning to fine-tune the model learned on a large code distance on a small code distance task often achieves better results.

[0069] Embodiment 1:

[0070] The embodiment of the present invention takes the performance of torus codes with different code lengths under various types of noise as an example to illustrate the feasibility and effect of the present invention. To this end, the present invention selects torus codes with code lengths of 5, 7, and 9, and considers bit-phase flip noise, symmetric and asymmetric depolarization noise models. Bit flip and phase flip noise are not considered separately here, because the MWPM algorithm is considered to be close to optimal in correcting such noise. Under each noise model, errors are randomly introduced into each data qubit with different probabilities, and then the corresponding error syndrome is extracted as a data set. Figure 6 It is a schematic diagram of the mapping relationship from quantum bits on the torus code to classical data provided by an embodiment of the present invention; the measurement results on the measurement quantum bits constitute the training data set of the low-level decoder (upper right corner), and the recovery operations on the data quantum bits constitute the data set labels of the low-level decoder (lower right corner).

[0071] Specifically, a Pauli error channel can be represented by an array (1-p, p x , p y , p z ), where each value corresponds to the probability of a Pauli error I (no error), X, Y, and Z. Therefore, the physical error rate on each data qubit is p = p x +p y +p z , this embodiment always considers the case of independent and identical distribution. The deviation η is defined as the ratio of the probability of Pauli Y error to the total probability of non-Y Pauli error, that is, At the same time, for the sake of simplicity, this embodiment only considers the special case p x =p z Then when the physical error rate is p, the probability of Y error occurring is The probability of X and Z errors occurring is Note that when η = 0.5, the probability of each Pauli error is p / 3, corresponding to a symmetric depolarization channel; when η → ∞, the probability of a Y error is p, corresponding to a bit-phase flip channel. For the error syndrome in the training dataset, when the MWPM decoder decodes successfully and the number of bits affected by the recovery operation is less than the number of bits in error, its label is set to the recovery chain predicted by the MWPM decoder, otherwise it is set to the set of original errors (which can also be regarded as a recovery chain operation).

[0072] The experimental results of this embodiment are as follows Fig. 9As shown, the horizontal axis is the physical error rate, the vertical axis is the decoding accuracy, and different curves represent the decoding performance of each decoder on torus codes of different sizes. It can be seen that the performance of the two-stage decoder SU-NetQD of the present invention is always better than that of the MWPM decoder under different physical error rates and code distances. As the error rate increases, the decoding accuracy of SQ-UNetQD and MWPM will decrease. However, the rate of decrease in the decoding accuracy of SU-NetQD of the present invention is much lower than that of MWPM, which reflects the stability of the performance of SU-NetQD. Under the conditions of the same error rate and noise type, the decoding accuracy of SU-NetQD of the present invention is much higher than that of MWPM. This is mainly because the neural network can learn the correlation between X and Z errors, and takes into account the error distribution under different physical error rates, so it can provide more accurate decoding results.

[0073] The curves of different distances under the same decoder intersect at one point. When the physical error rate is less than this threshold, a larger code distance will reduce the logical error rate, but when p is greater than the threshold, a larger code distance will lead to a higher logical error rate. This trend can be attributed to the increase in the number of errors introduced by the larger system size at high error rates, which masks the benefits of using more qubits to bring greater redundancy. Fig.10 A comparison chart of the change trend of the threshold measured by SU-NetQD and MWPM of the present invention with the deviation coefficient η is given, the horizontal axis represents the ratio of the Y type error rate to the X and Z type error rates, and the vertical axis represents the error threshold of the torus code measured by the decoder; the intersection of the quantum error correction decoding curves of different code distances can measure the threshold of the torus code's own anti-noise ability. In general, when the deviation coefficient η is larger, the threshold measured by SU-NetQD is larger. When η→∞, that is, under the bit phase flip channel, the threshold measured by the decoder of the present invention is improved from 0.095 to 0.231 compared with the result obtained by MWPM.

[0074] Embodiment 2:

[0075] The embodiment of the present invention takes the performance of the post-processing of the decoding result of MWPM by the advanced decoder as an example to illustrate the scalability and effect of the advanced decoder in the present invention. Under different noises, several advanced decoders are trained with MWPM as the low-level decoder, which can determine what kind of logical errors MWPM introduces in the decoding process, thereby improving the decoding accuracy. Specifically, for bit flip noise, this embodiment trains an advanced decoder based on the decoding result of MWPM and the corresponding logical errors at an error rate of p = 0.13. It can receive the recovery chain and original error syndrome obtained by MWPM and predict the probability of each logical error. During the test, if the probability of a certain logical error is greater than a preset threshold, it is considered that such a logical error has occurred and the corresponding logical operation needs to be performed, otherwise it is considered that it has not occurred. The final recovery chain operation is the exclusive OR of the recovery chain operation given by MWPM and the corresponding logical operation. For depolarization noise and bit-phase flip noise, we train at an error rate of p = 0.17, and other operations are similar to bit flip noise.

[0076] The final experimental results are shown in Table 1, where better_MWPM refers to the result obtained after post-processing the decoding result of MWPM using an advanced decoder. It can be seen that after post-processing by the advanced decoder, the decoding accuracy of MWPM is often greatly improved, and even under the bit flip noise model, the advanced decoder can also produce a certain improvement in the decoding accuracy of MWPM. This also further illustrates that the MWPM decoder is only close to the optimal under bit flip noise, not the theoretical optimal. These findings highlight the effectiveness of integrating advanced neural network strategies (such as the advanced decoder of the present invention) into traditional quantum error correction schemes, especially in complex noise environments, where traditional methods (such as MWPM) may be insufficient.

[0077] Table 1

[0078]

Claims

1. A quantum torus code decoder, characterized in that: Includes low-level decoder and high-level decoder; The low-level decoder adopts a self-attention U-net (SU-Net) model, which consists of a two-dimensional U-net structure for extracting information within the lattice, a three-dimensional self-attention block for extracting the dependency of different types of error syndromes between lattices, and a residual block; The advanced decoder reuses three SU-Net models. First, two SU-Net models are used to extract the original error syndrome and the context information of the recovery chain respectively. Then, the intermediate features are added to fuse the information of the two and input into another SU-Net model for final feature extraction. Finally, a global average pooling layer is used to average all elements in each channel and directly used for classification. The low-level decoder uses the error syndrome as input and predicts the recovery chain operation; the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical errors introduced by the low-level decoder.

2. A quantum torus code decoding method using the decoder as claimed in claim 1, characterized in that: The steps include: Step 1: Use the MWPM algorithm to decode the error syndrome generated by the collected torus code and the recovery operation that needs to be performed in a noisy environment, and correct the recovery operation predicted by MWPM by detecting the logical quantum state of the torus code; Step 2: Build low-level decoder and high-level decoder; The SU-Net used in low-level decoding consists of a 2D U-Net structure that extracts information within the lattice, and a 3D self-attention block and residual block that extracts dependencies between lattices for different types of error syndromes; The advanced decoder reuses three SU-Net models. First, two SU-Nets are used to extract the original error syndrome and the context information of the recovery chain respectively. Then, the intermediate features are added to fuse the information of the two and input into another SU-Net for the final feature extraction. Finally, a global average pooling layer is used to average all elements in each channel and directly used for classification. Step 3: Use the error syndrome in step 1 as the training dataset feature vector and the recovery chain operation as the dataset label, train the low-level decoder built in step 2 to make it converge, and record the recovery chain operation predicted by the low-level decoder for the last time, as well as the introduced logical errors; Step 4: Take the error syndrome together with the recovery chain operation predicted by the low-level decoder in step 3 as the feature vector of the training dataset, take the logical errors occurring on the torus code as the dataset label, and train the model of the high-level decoder constructed in step 2 to make it converge; Step 5: For new errors generated on the torus code under the same noise, the low-level decoder uses the extracted error syndrome as input and iteratively predicts the recovery chain operation, and the high-level decoder uses the recovery chain predicted by the low-level decoder and the original error syndrome as input and predicts the logical errors introduced by the low-level decoder; The final decoding is successful if and only if the recovery chain predicted by the low-level decoder can make the error syndrome completely disappear, and the logical error predicted by the high-level decoder is consistent with the joint effect of the recovery chain and the original error.

3. A quantum torus code decoding method according to claim 2, characterized in that: The recovery chain predicted by the low-level decoder once cannot completely eliminate the original error syndrome; an iterative decoding strategy is used to first use the currently predicted recovery chain to eliminate part of the error syndrome, and then the remaining error syndrome is input into the low-level decoder again for prediction, and this operation is repeated until the error syndrome disappears or the pre-set maximum number of predictions is reached.

4. A computer program, characterized in that The computer program enables a computer to execute the method according to any one of claims 2 to 3.

5. An electronic device, characterized in that: include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device executes the method as claimed in any one of claims 2 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 2 to 3 is implemented.

7. A chip, characterized in that: include: A processor, used to call and run a computer program from a memory, so that a device equipped with the chip executes a method as claimed in any one of claims 2 to 3.

8. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method as claimed in any one of claims 2 to 3 is implemented.

Images