A quantum error correction method and related device

Abstract

This invention relates to a quantum error correction method and related equipment. The method includes: acquiring discrete syndrome measurement sequences generated by a quantum processor based on surface codes over multiple consecutive error correction cycles; inputting the discrete syndrome measurement sequences into a causal convolutional neural network to output error prediction information for the next error correction cycle, the error prediction information including the probability distribution of qubit errors; determining a set of high-risk qubits for the next error correction cycle based on the probability distribution; pre-constructing a restricted decoding space for the next error correction cycle based on the set of high-risk qubits; in the next error correction cycle, preferentially performing error chain matching within the restricted decoding space; if no error chain satisfying the verification condition is found within the restricted decoding space, backtracking to the full search space for matching; and outputting error correction operation instructions based on the matching results. This method at least partially solves the problems of poor quantum error correction performance and low efficiency in related quantum error correction methods.

Description

Technical Field

[0001] This invention relates to the field of quantum computing fault-tolerant processing technology, and in particular to a quantum error correction method and related equipment. Background Technology

[0002] Quantum error correction is a key technology for realizing fault-tolerant quantum computing. Among various quantum error-correcting codes, surface codes have become the mainstream quantum error correction scheme due to their high fault tolerance threshold and the physical advantages of two-dimensional proximity coupling. Surface codes obtain syndrome information by periodically measuring the stabilizer of the auxiliary qubit, thereby inferring errors that have occurred on the data qubit and applying correction operations. Real-time decoding is the core challenge for the engineering application of surface codes. The decoding delay must be much smaller than the coherence time of the qubit; otherwise, errors will continue to propagate while waiting for decoding, leading to a sharp increase in the logic error rate.

[0003] Quantum error correction methods in related technologies typically treat the syndrome obtained in each error correction cycle as an independent event and solve the error chain by matching in the entire search space using a traditional decoder. However, the error rate is limited by the assumption of independent and identically distributed noise and fails to utilize the correlation of quantum noise in the time dimension, resulting in poor quantum error correction performance and low efficiency.

[0004] There is currently no effective solution to the aforementioned problems in the relevant technologies. Summary of the Invention

[0005] The present invention provides a quantum error correction method and related equipment, which at least partially solves the problems of poor quantum error correction effect and low efficiency in related technologies.

[0006] To address the above problems, one aspect of this invention provides a quantum error correction method, comprising: Acquire discrete syndrome measurement sequences generated by a quantum processor based on surface codes over multiple consecutive error correction cycles; The discrete syndrome measurement sequence is input into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of qubit errors. Based on the probability distribution, a set of high-risk qubits is determined for the next error correction cycle, and a restricted decoding space for the next error correction cycle is pre-constructed based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; In the next error correction cycle, the control decoder prioritizes error chain matching within the restricted decoding space; if no error chain satisfying the verification conditions is found within the restricted decoding space, the control decoder backtracks to the full search space for matching; and outputs error correction operation instructions based on the matching results.

[0007] In some embodiments, the causal convolutional neural network includes multiple levels of one-dimensional convolutional layers, and the receptive range of the convolutional kernels only covers the current and past time steps; wherein, The causal convolutional neural network is deployed in a field-programmable gate array to perform real-time inference on streamed discrete syndrome measurement sequences.

[0008] In some embodiments, the causal convolutional neural network includes multiple levels of dilated convolutional layers, with the dilation coefficient of each level increasing sequentially to expand the receptive field of the causal convolutional neural network to cover long-term noise correlation. The network structure of the causal convolutional neural network achieves temporal causality through the unidirectional coverage characteristic of the convolutional kernel's time axis. During the training phase of the causal convolutional neural network, the training samples consist of the historical syndrome window and the true error location labels of its subsequent error correction cycles. The network structure also restricts any predicted output to depend only on the current and historical syndrome inputs.

[0009] In some embodiments, determining the set of high-risk qubits for the next error correction cycle based on the probability distribution includes: Calculate the prediction confidence level of the probability distribution; If the predicted confidence level is greater than a preset confidence threshold, then the qubits whose probability values ​​exceed the preset probability threshold are marked as high-risk qubits; wherein, the preset confidence threshold is dynamically adjusted according to the decoding success rate of historical error correction cycles.

[0010] In some of these embodiments, it also includes: If the prediction confidence is less than or equal to the preset confidence threshold, the construction step of the restricted decoding space is skipped, and the decoder is directly controlled to solve the error chain matching problem in the full search space in the next error correction cycle.

[0011] In some embodiments, before inputting the discrete syndrome measurement sequence into the causal convolutional neural network, the method further includes: Obtain the physical location coordinates of each qubit on the quantum chip; The physical location coordinate information is fused with the discrete syndrome measurement sequence and used as a joint input to a causal convolutional neural network, so that the causal convolutional neural network can perceive non-uniform spatial noise distribution.

[0012] In some embodiments, the training steps of the causal convolutional neural network include: A training dataset is generated using a quantum error correction simulator; during the generation of the training dataset, random perturbations are applied to the physical bit error rate parameter to simulate quantum error correction environments under different noise intensities; Based on the physical bit error rate parameter after applying random perturbation, sample pairs containing discrete syndrome measurement sequences and corresponding true error locations are generated; The sample pairs are input into the causal convolutional neural network to be trained, and the difference between the predicted error probability distribution and the true error location is used as the loss function to update the parameters of the causal convolutional neural network.

[0013] To address the above problems, one aspect of the present invention provides a quantum error correction device, comprising: The acquisition module is used to acquire discrete syndrome measurement sequences generated by the quantum processor based on surface codes over multiple consecutive error correction cycles; The prediction module is used to input the discrete syndrome measurement sequence into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of qubit errors. A construction module is used to determine the set of high-risk qubits in the next error correction cycle based on the probability distribution, and to pre-construct a restricted decoding space for the next error correction cycle based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; The error correction module is used to control the decoder to prioritize error chain matching and solving within the restricted decoding space in the next error correction cycle; if no error chain satisfying the verification condition is matched within the restricted decoding space, the decoder is controlled to fall back to the full search space for matching and solving; and the error correction operation instruction is output according to the matching and solving result.

[0014] To address the aforementioned problems, one aspect of this invention provides a quantum computing error correction system, comprising: At least one quantum processor is used to execute a surface code error correction loop and generate a discrete syndrome measurement sequence; At least one of the above-described quantum error correction devices, wherein the quantum error correction device is coupled to the quantum processor; The quantum error correction device is equipped with a field-programmable gate array (FPGA), and the causal convolutional neural network is deployed in the logic resources of the FPGA.

[0015] To address the aforementioned problems, one aspect of this invention provides a non-transient machine-readable medium storing computer instructions for causing a computer to execute any of the aforementioned quantum error correction methods.

[0016] The beneficial effects of this invention are as follows: A discrete syndrome measurement sequence generated based on surface codes by an acquisition quantum processor within multiple consecutive error correction cycles is obtained; the discrete syndrome measurement sequence is input into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of qubit errors; based on the probability distribution, a set of high-risk qubits is determined for the next error correction cycle, and based on the set of high-risk qubits, a restricted decoding space for the next error correction cycle is pre-constructed; wherein, the restricted decoding space is a subset of the full search space; in the next error correction cycle, the decoder is controlled to preferentially decode within the restricted space. The invention employs a technique of matching and solving error chains within the decoding space; if no error chain satisfying the verification conditions is found within the limited decoding space, the decoder is controlled to backtrack to the full search space for matching and solving; and error correction operation instructions are output based on the matching and solving results. This technique overcomes the problems of poor quantum error correction effect and low efficiency in related quantum error correction methods. The embodiment of this invention obtains temporal information within multiple consecutive error correction cycles, introduces a causal convolutional neural network to extract temporal patterns, transforms these patterns into spatial constraints, and finally combines a priority search and backtracking search mechanism to achieve the technical effect of improving quantum error correction effect and efficiency.

[0017] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the main process of a quantum error correction method according to one embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of a causal convolutional neural network according to one embodiment of the present invention; Figure 3 This is a schematic diagram of the main modules of a quantum error correction device according to one embodiment of the present invention; Figure 4 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0020] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0021] Quantum error correction schemes in related technologies mainly fall into two categories in their surface code decoders: The first category is traditional decoders, such as the Minimum Weight Perfect Matching (MWPM) algorithm and the Union-Find algorithm. These decoders treat the syndrome obtained in each error correction cycle as an independent event and solve for the error chain by matching within the entire search space. While they can achieve microsecond-level decoding latency (e.g., the average latency for real-time decoding of the distance-5 surface code can reach 63 μs), their error rate is limited by the assumption of independent and identically distributed noise. They fail to utilize the temporal correlation of quantum noise—in real quantum systems, once a bit is erroneous, its subsequent error probability increases significantly due to error propagation, and environmental noise (such as fluctuations in a two-level system) exhibits temporal correlation, resulting in error patterns across multiple consecutive error correction cycles not being independent. Because they discard historical syndrome information, the logical error rate of traditional decoders is difficult to further reduce.

[0022] The second category is artificial intelligence decoders. For example, Google DeepMind's AlphaQubit, published in *Nature*, employs a Transformer architecture and achieves a 6% reduction in error rate compared to tensor network methods and a 30% reduction compared to correlation-matching decoders. However, AlphaQubit's inference latency is approximately 100 ms, which cannot meet the real-time requirements of megahertz-level syndrome detection in superconducting quantum processors. Furthermore, the quadratic complexity of the Transformer makes it difficult to deploy in a lightweight manner on low-latency hardware platforms such as field-programmable gate arrays.

[0023] In summary, the quantum error correction methods in related technologies have the following technical defects: traditional decoders ignore the causal relationship of the syndrome time series and treat each error correction as an independent event, resulting in limited decoding accuracy and poor quantum error correction effect; high-precision AI decoders have too large inference latency and cannot run in real time, resulting in low quantum error correction efficiency.

[0024] To address the aforementioned problems, embodiments of the present invention provide a quantum error correction method, such as... Figure 1As shown, this quantum error correction method mainly includes: Step S101: Obtain the discrete syndrome measurement sequence generated by the quantum processor based on the surface code in multiple consecutive error correction cycles; Step S102: Input the discrete syndrome measurement sequence into the causal convolutional neural network to output the error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of the qubits causing errors; Step S103: Based on the probability distribution, determine the set of high-risk qubits in the next error correction cycle, and based on the set of high-risk qubits, pre-construct a restricted decoding space for the next error correction cycle; wherein, the restricted decoding space is a subset of the full search space; Step S104: In the next error correction cycle, control the decoder to prioritize error chain matching and solving within the limited decoding space; if no error chain that meets the verification conditions is matched within the limited decoding space, control the decoder to backtrack to the full search space for matching and solving; output error correction operation instructions based on the matching and solving results.

[0025] Based on the above setup, by acquiring temporal information within multiple consecutive error correction cycles, a causal convolutional neural network is introduced to extract temporal patterns. These patterns are then transformed into spatial constraints. Finally, combined with priority search and backoff search mechanisms, the technical effects of improving quantum error correction performance and efficiency are achieved. Specifically, by utilizing the temporal correlation of historical syndromes, the causal convolutional neural network can predict the trend of error occurrence, focusing the decoder's attention on truly high-risk regions where errors are likely to occur. Compared to traditional decoders that do not utilize temporal information, this embodiment of the invention reduces the probability of missed and false detections, thereby reducing the logical error rate and improving quantum error correction accuracy. By pre-constructing a constrained decoding space, the decoder's search range is significantly reduced from the entire space, resulting in a significant decrease in the time consumption of error chain matching and solving, thus improving quantum error correction efficiency. Simultaneously, the backoff mechanism ensures that decoding can still be completed in the worst-case scenario (model prediction failure), preventing indefinite stagnation due to prediction errors and enhancing the robustness of the quantum error correction system.

[0026] Specifically, noise in quantum systems is not time-independent white noise. In real-world scenarios, when a qubit makes an error, the probability of that qubit inducing errors in adjacent qubits in subsequent cycles increases significantly due to the error propagation effect of two-qubit gate operations. Simultaneously, low-frequency noise sources in the environment (such as charge noise and magnetic flux noise) remain correlated over multiple consecutive cycles. Traditional decoders only use single-cycle syndromes, essentially discarding this historical information containing predictive value. This embodiment of the invention, based on step S101 above, collects sequences from multiple consecutive cycles (because temporal correlations can only be observed and modeled within a sequence; a single point cannot reflect a trend, while multiple points constitute a predictable temporal pattern), expanding the decoder's input from a syndrome of a single error correction cycle to a sequence of multiple consecutive cycles, enabling subsequent processing to utilize the temporal correlation of quantum noise.

[0027] According to an embodiment of the present invention, based on the above step S102, a causal convolutional neural network is used to perform time-series modeling of the historical syndrome sequence, outputting the error probability distribution of each qubit in the next cycle. The "causal" characteristic of the causal convolutional neural network—the convolution kernel only covers the current and past time steps—ensures that the prediction process does not use future information, conforming to the physical constraints of real-time decoding. Compared to the posterior matching of traditional decoders, the above steps achieve prior prediction of impending errors, transforming error correction from a passive response to an active prediction. Furthermore, compared to the Transformer's self-attention mechanism (complexity O(n²)), the causal convolutional neural network has a small parameter count and low inference latency, and can be deployed on a Field Programmable Gate Array (FPGA) for real-time inference.

[0028] In some embodiments, based on step S103 above, the probability distribution is transformed into an operable spatial constraint (i.e., a set of high-risk qubits), and a restricted decoding space much smaller than the full search space (which contains all possible combinations of error positions, but actual physical noise causes most bits to have an extremely low error probability) is defined accordingly. This transformation achieves two key functions: first, it hardens the soft prediction results (probability distribution) of the neural network model into hard constraints (spatial boundaries) that the decoder can directly utilize; second, it significantly reduces the search range of the decoder from the full space to the high-risk region. Since the decoding complexity of quantum error correction is positively correlated with the size of the search space, this directly reduces the computational overhead and solution time of the decoder.

[0029] In some examples, based on step S104 above, firstly, the decoder prioritizes solving within the limited decoding space. Due to the reduced space, the matching speed is significantly improved, meeting the latency requirements of real-time decoding (e.g., less than 50 μs). Secondly, when an error chain satisfying the verification conditions cannot be found within the limited space (i.e., when the causal convolutional neural network model predicts an error or noise anomaly), the system automatically backtracks to the full search space, ensuring that the error correction performance is no less than that of a traditional decoder. This dual solution mechanism of "priority + backtracking" balances the efficiency improvement brought by the accelerated prediction of the neural network model with the robustness guarantee of traditional methods, avoiding the catastrophic failure that may result from relying solely on the prediction of the neural network model. The reason why "prioritizing error chain matching within the limited decoding space" can improve efficiency is that the decoder operates on a smaller matching graph or subgraph, reducing the number of nodes and edges involved in the matching, thereby reducing the required computation, storage access, and solution latency. The reason why "backing back to the full search space for matching and solving" can guarantee robustness is that when the neural network model makes a prediction error (the real error is not in the restricted decoding space) or the confidence of the neural network model is too low (the input noise pattern exceeds the training distribution), the error chain that meets the verification conditions cannot be found in the restricted decoding space (that is, the residual after matching does not meet the stabilizer constraint). In this case, backing back to the full space is equivalent to calling the traditional decoder as a safety net.

[0030] It is also understandable that, since the verification condition (the matching result must be consistent with the actual syndrome) is an objective physical constraint and is not affected by the prediction of the neural network model, the backoff trigger condition is objectively determinable, rather than relying on the confidence estimate of the prediction result of the neural network model, thus avoiding the risk of decoding failure due to false belief and incorrect prediction.

[0031] In some of these embodiments, the causal convolutional neural network includes multiple levels of one-dimensional convolutional layers, and the sensing range of the convolutional kernels only covers the current and past time steps; wherein the causal convolutional neural network is deployed in a field-programmable gate array to perform real-time inference on streaming discrete syndrome measurement sequences.

[0032] Based on the above setup, causal guarantees are provided at the algorithm level through causal convolutional neural networks, low latency guarantees are provided at the hardware level through FPGAs, and real-time guarantees are provided at the system level through streaming processing. This allows the prediction results of the neural network model to be truly embedded into the real-time control loop of quantum error correction, satisfying both physical constraints and achieving engineering-usable latency metrics.

[0033] Specifically, the real-time requirement of quantum error correction necessitates that the decoder must complete the processing of the current cycle before the start of the next error correction cycle and cannot predict future syndromes. Simultaneously, in temporal data processing, the sliding window direction of the convolution kernel determines the direction of information flow. Based on these settings, by limiting the perception range of the convolution kernel in a causal convolutional neural network to include only the inputs of the current time step and previous time steps in the convolution calculation at each time step, and masking inputs after the current time step, the output of each layer of the causal convolutional neural network is strictly independent of information from future time steps. This structural causal constraint of the neural network model does not require additional masking or post-processing, guaranteeing the temporal causality of the predictions from the network architecture itself. This aligns with the physical requirement in quantum error correction that "real-time decoding cannot predict future syndromes," thus satisfying the hard physical requirement for causality. Furthermore, after stacking multiple levels of one-dimensional convolutional layers, the receptive field of the higher-level convolutional layers expands linearly, enabling causal convolutional neural networks to gradually expand from short-range dependencies of a few cycles to long-range dependencies of dozens of cycles. This allows them to simultaneously capture rapid fluctuations (such as single erroneous impulses) and slow-changing trends (such as environmental noise drift). Different layers automatically learn features at different time scales, with shallow layers focusing on local mutations and deep layers focusing on overall patterns. This allows them to capture dependencies at different time scales without the need for manually designing time windows.

[0034] Meanwhile, Field-Programmable Gate Arrays (FPGAs) are hardware-reconfigurable devices whose computational logic is executed directly as hardware circuits, rather than through an instruction pipeline. The aforementioned causal convolutional neural networks are characterized by: regular computational patterns (one-dimensional sliding convolution), linear data flow (sequential input at time steps), and fixed weights (unchanged after pre-training). These characteristics are precisely what FPGAs excel at—numerous parallel multiply-accumulate operations can be mapped to hardware multipliers and adder trees; weights can be stored in on-chip registers or block random access memory, eliminating the need for frequent reads from external memory; and the data flow can be pipelined to output a result per clock cycle. By deploying causal convolutional neural networks within FPGAs, the algorithmic structure is transformed into a hardware advantage, enabling causal convolutional networks to complete inference with μs-level latency. This helps further improve quantum error correction efficiency and can meet the real-time error correction requirements of superconducting qubits.

[0035] Furthermore, in traditional methods, neural network inference typically requires operating system intervention (such as driver calls and interrupt handling), introducing latency jitter at the level of hundreds of microseconds or even milliseconds. In this embodiment of the invention, when performing real-time inference on discrete syndrome measurement sequences based on FPGA with streaming input (streaming input means that the syndrome data generated in each error correction cycle is immediately sent to the network after generation, rather than waiting to collect the complete batch before processing), the entire network is directly integrated into the quantum electronics control system as part of the hardware logic. Syndrome data is sent in through direct memory access or a dedicated interface, without going through the operating system network stack. This is equivalent to changing the neural network model prediction from "software function calls" to "hardware data streams," eliminating uncertain latency at the software level.

[0036] In some embodiments, the causal convolutional neural network described above includes multiple levels of dilated convolutional layers, with the dilation coefficient of each level increasing sequentially to expand the receptive field of the causal convolutional neural network to cover long-term noise correlation.

[0037] Based on the above settings, by leveraging three specific features of causal convolutional neural networks—multi-level dilated convolutional layers, sequentially increasing dilation coefficients, and expanding the receptive field to cover long-period noise—the predictive ability of the neural network model for long-term noise patterns is significantly improved without increasing inference latency or the number of parameters.

[0038] First, the mathematical properties of dilated convolution provide a computational match for the long-range correlation of quantum noise. In quantum error correction, the temporal correlation of noise manifests as follows: the error probability of the current correction cycle depends not only on the previous correction cycle (due to insufficient Markov property) but also on earlier states (such as the memory effect of charge trapping / releasing processes). To predict the error of the next correction cycle, the neural network needs access to a sufficiently long history. Ordinary convolution requires stacking a large number of layers to achieve a long receptive field, but each additional layer introduces nonlinear activation and extra computation, leading to a significant increase in pipeline latency on FPGAs. Multi-level dilated convolution expands the receptive field exponentially, allowing the history of dozens of cycles to be covered with only a few layers (e.g., 5 layers), satisfying the need for long-range dependence while controlling the computational depth.

[0039] Secondly, the sequentially increasing expansion coefficient constructs a hierarchical temporal feature extractor. The temporal spectrum of quantum noise is typically complex: it contains high-frequency components (such as random bit flips) and low-frequency components (such as drift caused by ambient temperature fluctuations). Single-scale neural network models can only learn patterns within a certain frequency band, limiting their predictive ability across the entire spectrum. In contrast, networks with increasing expansion coefficients learn high-frequency, rapidly changing patterns at shallower layers and low-frequency, slow-moving trends at deeper layers, enabling automatic multi-resolution analysis. This structure naturally aligns with the physical characteristics of quantum noise—noise from different physical sources has different time constants. Causal convolutional neural networks do not need to know these constants beforehand; instead, they adaptively learn them through the increasing expansion coefficient. This results in the probability distribution output by the causal convolutional neural network integrating information from various time scales, making it more accurate than single-scale models.

[0040] Furthermore, expanding the receptive field does not lead to overfitting; on the contrary, it improves generalization ability. Generally, a larger receptive field means more parameters and stronger fitting ability, making it prone to overfitting training noise. However, the temporal correlation of quantum noise is a universal characteristic of physical systems, not a special case of the training set. When a causal convolutional neural network learns this universal long-range pattern, it can make robust predictions based on long-range patterns for unseen noise intensities or slightly varying noise environments. For example, even if the physical error rate increases by 50% during testing, the temporal correlation structure of the noise (such as the relaxation time of a noise source) remains unchanged, and the causal convolutional neural network remains effective. Therefore, based on this feature, while improving the prediction accuracy of the neural network model, it also enhances the model's robustness to changes in noise levels.

[0041] In some of these embodiments, the network structure of the causal convolutional neural network achieves temporal causality through the unidirectional coverage characteristic of the convolutional kernel's time axis. During the training phase of the causal convolutional neural network, the training samples consist of the historical syndrome window and the true error location labels of its subsequent error correction cycles. The network structure restricts any predicted output to depend only on the current and historical syndrome inputs.

[0042] Based on the above settings, the causality during inference is guaranteed by structural constraints on the causal convolutional neural network, and the causal mapping of the learning process is guaranteed by the temporal construction of training samples. This achieves strict guarantee of the temporal causal consistency of the neural network model throughout the entire training and inference stages, enabling the neural network model to be deployed without bias in a real-time streaming environment.

[0043] Specifically, according to embodiments of the present invention, on one hand, the convolution kernel of the causal convolutional neural network is designed to have a unidirectional coverage, meaning that at each time step, the convolution kernel only performs a weighted summation of the inputs at the current time step and several previous time steps, excluding any inputs after the current time step. This unidirectional coverage characteristic in the structure guarantees from the root of the computation graph that the value at any output position is determined only by that position and earlier inputs. For quantum error correction, real-time decoding requires completing the decoding of the current cycle before the syndrome of the next cycle has been measured; therefore, the neural network model absolutely cannot rely on future information. This feature enforces this constraint through the physical design of the network structure (rather than post-processing masks), eliminating any possible information leakage paths.

[0044] On the other hand, during the training phase, each training sample consists of a historical syndrome window and the true error location labels of its subsequent error correction cycles. Since the network structure itself guarantees that no future information is introduced into the forward computation, the model can learn the correct causal mapping from the past to the future simply by using a standard supervised learning process (input historical window, output future label). This training method avoids any shortcuts that rely on future information, ensuring that the model can make accurate predictions based solely on historical input during real-time inference.

[0045] Furthermore, quantum processors typically generate syndrome data at fixed intervals (usually on the microsecond scale), and the decoder must complete all processing for the current cycle before the next cycle begins. If the aforementioned causal convolutional neural network learns future information during training, then in real-time operation, since the future information has not yet been generated, the neural network model can only make suboptimal predictions, resulting in an actual error correction rate lower than the training metrics. Because the causal convolutional neural network in this scheme is structurally guaranteed to never rely on future information during inference, the evaluation metrics on the training set can accurately reflect the performance during real-time deployment, avoiding the trap of "good performance during training but failure during runtime." This further enhances the practicality of the quantum error correction method provided by this invention in real-world engineering applications.

[0046] According to a specific embodiment of the present invention, such as Figure 2The diagram illustrates the specific network topology of the causal convolutional neural network provided in this embodiment of the invention. This causal convolutional neural network is designed as a prediction engine. Its main function is to receive the historical syndrome data stream (i.e., the aforementioned discrete syndrome measurement sequence) generated by the quantum processor, extract spatiotemporal features through multi-layer convolution operations, and finally output the probability prediction of errors in each data bit at the next time step. This architecture exhibits significant unidirectional temporal causality, meaning that the network's computation relies only on current and past information, strictly prohibiting the use of information from future time steps, thereby ensuring low-latency streaming inference in real-time quantum error correction loops. Specifically, the network topology of this causal convolutional neural network includes: (1) Input layer (used to input the historical syndrome sequence matrix generated by the quantum processor) Input data: The historical syndrome measurement sequence input to the causal convolutional neural network model is a two-dimensional time series matrix with dimensions defined as 24 (space) × 20 (time); where, the spatial dimension (24): corresponds to the 24 stable syndrome measurement channels in the surface code at a distance of d=5, or the stable measurement results measured by 24 auxiliary qubits; the temporal dimension (20): corresponds to the past 20 consecutive error correction cycles "observed" by the network under the sliding window mechanism.

[0047] The aforementioned input layer transforms the dynamic noise evolution process of the quantum processor into a digital signal that can be processed by the neural network, capturing the temporal and spatial correlation of the noise.

[0048] (2) Feature extraction layer: contains three cascaded one-dimensional convolutional layers.

[0049] The first stage is a causal convolutional layer. Its parameters are configured as follows: kernel size k=3, dilation coefficient d=1, and number of output channels C=32. As the entry point of the network, the causal convolutional layer directly extracts local features from the original input sequence. Since d=1, this causal convolutional layer only focuses on the local correlation between adjacent time steps, establishing the basic temporal causal chain.

[0050] The second and third stages are dilated convolutional layers. Their parameter configurations are as follows: the dilation coefficient d=2 and the number of output channels C=64 for the second stage; and the dilation coefficient d=4 and the number of output channels C=64 for the third stage. By inserting gaps between the convolutional kernel elements (controlled by the dilation coefficient d), the causal convolutional neural network can exponentially expand its receptive field without increasing the number of parameters or computational complexity. Simultaneously, since quantum noise often exhibits long-range temporal correlation (e.g., low-frequency noise or drift), by using progressively increasing dilation coefficients (1→2→4), the causal convolutional neural network can capture historical noise patterns over longer spans, thereby more accurately predicting future erroneous trends.

[0051] (3) Output layer (used for fully connected mapping and probability prediction) The fully connected mapping layer is implemented using either one-dimensional convolution (Conv1D) or a fully connected layer (FC), with 16 output channels and a sigmoid activation function. This layer maps abstract features extracted from higher layers back to the specific physical bit space. Its output is a prediction of the error probability of the data bit, formatted as a 25-dimensional vector. Each element of the vector corresponds to the probability value (range 0-1) of a data bit occurring in the next error correction cycle. The sigmoid function ensures that the output value is normalized to a probability distribution, directly serving the subsequent determination of high-risk qubit sets.

[0052] Meanwhile, in the aforementioned causal convolutional neural network, the convolutional kernels only cover the current and past time steps. The neural network model does not need to wait for future data during inference, eliminating the latency introduced by traditional bidirectional networks. Furthermore, the computational patterns of one-dimensional convolution and dilated convolution are highly regular, making them ideal for deployment within the hardware logic resources of FPGAs. Moreover, this architecture allows data to enter the chip in a streaming manner and complete inference within extremely short clock cycles, meeting the stringent requirements of quantum error correction for microsecond or even nanosecond-level response speeds.

[0053] In some embodiments, the step of determining the set of high-risk qubits in the next error correction cycle based on the probability distribution includes: calculating the prediction confidence of the probability distribution; if the prediction confidence is greater than a preset confidence threshold, then marking the qubits whose probability values ​​exceed the preset probability threshold as high-risk qubits; wherein the preset confidence threshold is dynamically adjusted according to the decoding success rate of historical error correction cycles.

[0054] Based on the above settings, a specific implementation method for determining the set of high-risk qubits based on probability distribution is provided. It constructs a decision gate for whether to adopt the prediction results of the causal convolutional neural network through a two-layer judgment of a pre-set confidence threshold and a preset probability threshold. At the same time, the dynamic adjustment mechanism based on the preset probability threshold enables the system to automatically maintain optimal performance under various noise levels without the need for manual retraining or manual parameter tuning.

[0055] The probability distribution output by a causal convolutional neural network is a high-dimensional vector representing the error probability of each qubit. However, for the same set of inputs, a causal convolutional neural network sometimes outputs a sharp probability distribution (a few qubits have probabilities close to 1, and the rest close to 0), and sometimes it outputs a flat probability distribution (all qubits have probabilities of approximately 0.5). The prediction confidence introduced in this embodiment is a scalar that measures the determinism of the neural network model by aggregating information from the probability distribution (e.g., taking the difference between the highest and second-highest probabilities, or calculating the entropy of the distribution). The prediction confidence value is high when the neural network model is highly certain about the prediction result; and low when the neural network model is uncertain. Therefore, by calculating the prediction confidence, subsequent steps can make adaptive decisions based on the intrinsic state of the neural network model, rather than blindly trusting its output.

[0056] Specifically, according to one embodiment of the present invention, a fixed confidence threshold is first preset to distinguish between "credible predictions" and "uncredible predictions". When the prediction confidence corresponding to the probability distribution output by the causal convolutional neural network model exceeds the preset confidence threshold, it indicates that the currently input historical syndrome sequence conforms to the noise pattern learned by the neural network model, and the probability distribution given by the neural network model is relatively reliable. At this time, a second threshold (preset probability threshold) is applied to filter individual bits, and bits with probability values ​​exceeding the preset probability threshold are marked as high-risk qubits. Based on this two-layer screening mechanism: the first layer of screening (screening according to the preset confidence threshold) excludes overall uncredible predictions, and the second layer of screening (screening according to the preset probability threshold) extracts the most likely erroneous key bits (i.e., high-risk qubits) from credible predictions. The two thresholds have a clear division of labor: the preset confidence threshold controls whether the prediction mode is enabled, and the preset probability threshold controls the sensitivity of the prediction mode. It can be understood that, compared to directly outputting all bits with probabilities exceeding the preset probability threshold (regardless of whether the model is confident), this feature avoids outputting false high-risk bits when the causal convolutional neural network model is unreliable, thereby reducing the mislabeling rate.

[0057] Furthermore, the noise environment of a quantum system may change slowly over time (e.g., due to device aging, temperature fluctuations, or changes in electromagnetic interference) or change abruptly (e.g., a sudden increase in noise caused by a cosmic ray strike). If the preset confidence threshold remains fixed, when the environmental noise characteristics deviate from the training distribution, the confidence of the neural network model may systematically decrease, causing the prediction mode to be frequently skipped, and the system to degenerate into full-space decoding, losing the accuracy advantage brought by model prediction. Conversely, if the noise environment becomes more stable than during training, the confidence of the neural network model may be systematically higher, but the fixed preset confidence threshold may still enable the prediction mode, while the actual prediction accuracy of the neural network model may have decreased. According to a specific embodiment of the present invention, the confidence threshold can be dynamically adjusted by monitoring the actual decoding success rate (i.e., the proportion of successful matches within the limited decoding space or the logic error rate after error correction). Specifically, when the recent decoding success rate is higher than the target value, the preset confidence threshold is appropriately lowered to allow more predictions to be adopted; when the decoding success rate is lower than the target value, the preset confidence threshold is increased to make the system more cautious. Based on the aforementioned adaptive adjustment strategy for the preset confidence threshold, the system always operates at the optimal "prediction-backoff" balance point, requiring no manual intervention and adapting to environmental changes.

[0058] In some embodiments, the method further includes: if the prediction confidence is less than or equal to a preset confidence threshold, skipping the construction step of the restricted decoding space and directly controlling the decoder to solve the error chain matching problem in the full search space in the next error correction cycle.

[0059] Based on the above settings, when the prediction of the neural network model is unreliable, it is possible to switch to the traditional decoding path with guaranteed performance at minimal cost, thereby maintaining the stability of the overall decoding latency and success rate of the system.

[0060] According to an embodiment of the present invention, the construction of a restricted decoding space requires determining a set of high-risk qubits based on the probability distribution output by a neural network model, and then using graph theory to determine a subgraph range (e.g., a connected region containing the high-risk qubits and their neighboring qubits). The process of determining the set of high-risk qubits involves traversing candidate nodes, calculating adjacency relationships, and allocating memory, all of which consume clock cycles and logic resources on either an FPGA or CPU. When the prediction confidence corresponding to the output of the neural network model is low, it means that the current input historical syndrome sequence deviates significantly from the training distribution, indicating insufficient reliability of the probability distribution output by the surface neural network model. In this case, the set of high-risk qubits constructed based on this probability distribution is likely to not contain the true error locations, resulting in an invalid or inefficient restricted decoding space. The decoder is highly likely to fail in its search within this space, forcing a regression to the full space. To address this situation, the above-described skipping of the restricted decoding space construction step, based on an embodiment of the present invention, directly avoids this inevitable computational overhead, effectively pruning a low-quality execution path at the decision-making level.

[0061] Furthermore, the full search space is the standard input range for surface code decoding. Traditional decoders (such as minimum weight perfect matching or joint search algorithms) can guarantee finding the error chain that meets the verification conditions within the full search space (as long as the noise rate is below the error correction threshold). When the prediction confidence corresponding to the output of the neural network model is low, directly calling full space decoding is equivalent to bypassing model prediction and switching to a traditional decoding process without prediction assistance. Since traditional decoders do not rely on any historical information or probability prediction, their performance lower limit is known and acceptable. It is understandable that if an invalid or inefficient restricted decoding space is constructed by forcibly using the neural network model prediction and then backtracking after failure, it will not only consume more time for construction and failure search, but may also affect subsequent cycles due to time constraints during backtracking. Full search space decoding is an alternative with deterministic performance guarantees. It provides a safe basis for model-assisted solutions. If the prediction confidence is less than or equal to a preset confidence threshold, the construction step of the restricted decoding space is skipped, and the decoder is directly controlled to solve the error chain matching in the full search space in the next error correction cycle. This avoids the path of "failure first and then remedy", so that the system can still maintain stable and predictable decoding latency and success rate in scenarios where the neural network model is not applicable.

[0062] In some embodiments, before inputting the discrete syndrome measurement sequence into the causal convolutional neural network, the method further includes: obtaining the physical position coordinate information of each qubit on the quantum chip; and performing feature fusion of the physical position coordinate information and the discrete syndrome measurement sequence as a joint input to the causal convolutional neural network, so that the neural network model can perceive the non-uniform spatial noise distribution.

[0063] Based on the above settings, the spatial heterogeneity prior of quantum chips is incorporated into the causal convolutional neural network model, enabling the neural network model to dynamically adjust its attention to different physical locations in time series prediction, thereby more accurately predicting the location of errors in non-uniform noise environments.

[0064] Specifically, qubits on a quantum chip are not homogeneous. Due to factors such as manufacturing process variations, wiring differences, and uneven distribution of nearby heat sources, qubits at different physical locations have different relaxation times, decoherence rates, and gate operation fidelity. These differences typically manifest as spatial correlations: for example, the temperature in the central region of the chip is higher, leading to an overall higher error rate for qubits in that region; edge qubits are more strongly coupled to environmental noise. The aforementioned physical location coordinate information (such as (x,y) coordinates on a two-dimensional plane) is an explicit numerical representation of this spatial heterogeneity. By feeding this physical location coordinate information as an additional input feature into a causal convolutional neural network model, the model can learn the mapping relationship between coordinates and error rates from the training data. For example, it might discover that qubits near coordinates (3,5) frequently make errors in historical data, thus assigning a higher prior probability to qubits located in that region during the prediction phase.

[0065] In this context, the discrete syndrome measurement sequence is data in the time dimension, while the physical location coordinates are static data in the spatial dimension. These two types of data differ structurally: the discrete syndrome measurement sequence (i.e., the time series) is dynamic and periodically updated, while the physical location coordinates are static and unchanging. According to a specific embodiment of the present invention, in the above feature fusion step, the physical location coordinates are first mapped to the same dimensional space as the discrete syndrome measurement sequence (e.g., by encoding the coordinates of each bit into a feature vector through an embedding layer and then extending it to all time steps). Then, the two are concatenated or added along the feature channels. The fused joint input simultaneously contains "the change in the measurement value of a certain stabilizer over time" and "the geometric position of the qubits covered by that stabilizer on the chip." This allows the causal convolutional neural network to perceive the physical region corresponding to the syndrome when processing the syndrome at each time step. For example, when the causal convolutional neural network sees an abnormal pattern in the syndrome from the lower left corner of the quantum chip, it indicates that the error rate in that region has historically been high, thus it tends to attribute the anomaly to local noise rather than global fluctuations.

[0066] According to embodiments of the present invention, noise distribution on actual quantum chips is non-uniform—the error rate in some "hotspot" regions may be several times higher than in other regions. Traditional models (which only input the syndrome sequence) can only indirectly infer this non-uniformity through historical data: for example, if a bit frequently errs in a historical period, the model can learn from the time series that the error probability of that bit is high. However, this learning is data-driven and requires a large number of error samples of that bit to support it. For newly calibrated chips or after the replacement of faulty bits, historical data may be scarce, making it difficult for the model to adapt quickly.

[0067] The sliding window operation of convolutional neural networks inherently possesses locality; injecting coordinate information is equivalent to adding a spatial label to each receptive field region, allowing the network to learn "spatial-temporal joint features." By directly inputting physical location coordinates during the training of a causal convolutional neural network model, the model can transfer and learn spatial patterns (even if a bit has few historical error samples, but its coordinates are located in a known hotspot region (e.g., the chip center), the model can still infer a high probability of error because it has already learned the overall mapping function between "coordinates and error rate" in the training data). Furthermore, for the ambiguity problem of different bits but the same syndrome pattern (two bits at different locations may produce the same stable submeasurement results, but their physical error rates are different), by fusing physical location coordinate information with discrete syndrome measurement sequences as joint input to the causal convolutional neural network, the model can make more accurate distinctions regarding such ambiguities.

[0068] In some embodiments, the training steps of the causal convolutional neural network include: generating a training dataset using a quantum error correction simulator; applying random perturbations to the physical bit error rate parameters during the generation of the training dataset to simulate quantum error correction environments under different noise intensities; generating sample pairs containing discrete syndrome measurement sequences and corresponding real error locations based on the physical bit error rate parameters after applying random perturbations; and inputting the sample pairs into the causal convolutional neural network to be trained, using the difference between the predicted error probability distribution and the real error location as the loss function to update the parameters of the causal convolutional neural network.

[0069] Based on the above settings, a causal convolutional neural network robust to fluctuations in physical error rate is trained by a simulated data generation strategy with random perturbations of noise rate, enabling it to maintain high prediction accuracy in the non-steady-state noise environment of real quantum systems.

[0070] Training neural network models on practical quantum hardware faces significant challenges. Each fine-tuning of the model requires numerous quantum operations, resulting in high time consumption and costs. Furthermore, labeling the true error locations necessitates complete state tomography, which is virtually impractical in real-time systems. Quantum error correction simulators (such as Stim) simulate the circuit-level noise and syndrome measurement process of surface codes on classical computers, generating syndrome sequences of arbitrary size in a short time, and inherently containing correct labels (the simulator knows where the errors occur). This invention utilizes a quantum error correction simulator, enabling offline batch generation of training data without consuming expensive quantum processor time, and avoiding the impact of physical device instability and noise drift on data consistency.

[0071] According to embodiments of the present invention, the physical bit error rate of a real quantum processor is not fixed, but fluctuates over time (e.g., due to temperature changes, laser power drift) or varies in different regions of the chip. If only a single error rate parameter is used to generate data during training, the neural network model will overfit to that specific noise level, and the prediction performance will significantly decrease when the error rate in the deployment environment is too high or too low. Based on the above embodiments of the present invention, when generating each sample, a random perturbation is added to the error rate parameter (e.g., multiplying it by a uniform random factor of [0.8, 1.2] around the original value), so that the training data covers a continuous spectrum from low noise to high noise. During training, the neural network model is forced to learn feature representations that are effective under different noise intensities, thereby maintaining a high prediction accuracy even when encountering precise error rate values ​​that did not appear in training during testing.

[0072] Specifically, the error propagation mechanism of surface codes has a smooth dependence on the noise rate, and the logical error rate is a continuous convex function of the physical error rate. Therefore, the range covered by random perturbation is sufficient for the neural network model to learn interpolation capabilities without having to sample each possible error rate value separately.

[0073] In some embodiments, supervised learning training of causal convolutional neural networks requires extensive pairing of inputs (multi-cycle syndrome sequences) and outputs (true error locations in the next cycle). In a simulator, the error rate parameter is randomly perturbed, and then a simulation of the error correction cycle for the surface code is performed to obtain a continuous syndrome sequence. Since the simulator internally knows the error location injected at each step, it can accurately label the accumulated errors (i.e., the targets to be corrected) on each data bit at the end of each error correction cycle. In this embodiment of the invention, by utilizing the inherent observability of the simulator, aligned sequences and labels are automatically generated without manual annotation or post-processing. The simulator is a deterministic or pseudo-random simulation, and all random number seeds and error injection records for each run are traceable. This ensures that the annotation process does not introduce additional noise, and the labels are strictly time-aligned with the inputs (information from future time steps is explicitly avoided).

[0074] The quantum error correction method provided in this invention employs the following techniques: acquiring discrete syndrome measurement sequences generated by a quantum processor based on surface codes over multiple consecutive error correction cycles; inputting these discrete syndrome measurement sequences into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein the error prediction information includes the probability distribution of qubit errors; determining a set of high-risk qubits for the next error correction cycle based on the probability distribution; and pre-constructing a restricted decoding space for the next error correction cycle based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; in the next error correction cycle, controlling the decoder to preferentially perform error chain matching within the restricted decoding space; if no error chain satisfying the verification conditions is found within the restricted decoding space, controlling the decoder to backtrack to the full search space for matching; and outputting error correction operation instructions based on the matching results. By acquiring temporal information over multiple consecutive error correction cycles, introducing a causal convolutional neural network to extract temporal patterns, transforming these temporal patterns into spatial constraints, and finally combining a priority search and backtracking search mechanism, the technical effect of improving quantum error correction performance and efficiency is achieved.

[0075] Based on the quantum error correction method provided in the embodiments of the present invention, the present invention also provides a quantum error correction device. Figure 3 As shown, the quantum error correction device 300 includes: The acquisition module 301 is used to acquire the discrete syndrome measurement sequence generated by the quantum processor based on the surface code in multiple consecutive error correction cycles; The prediction module 302 is used to input the discrete syndrome measurement sequence into the causal convolutional neural network to output the error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of the qubits causing errors; Module 303 is used to determine the set of high-risk qubits in the next error correction cycle based on the probability distribution, and to pre-construct a restricted decoding space for the next error correction cycle based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; The error correction module 304 is used to control the decoder to prioritize error chain matching and solving within the limited decoding space in the next error correction cycle; if no error chain that meets the verification conditions is matched within the limited decoding space, the decoder is controlled to fall back to the full search space for matching and solving; and the error correction operation command is output according to the matching and solving result.

[0076] Based on the above setup, by acquiring temporal information within multiple consecutive error correction cycles, a causal convolutional neural network is introduced to extract temporal patterns. These patterns are then transformed into spatial constraints. Finally, combined with priority search and backoff search mechanisms, the technical effects of improving quantum error correction performance and efficiency are achieved. Specifically, by utilizing the temporal correlation of historical syndromes, the causal convolutional neural network can predict the trend of error occurrence, focusing the decoder's attention on truly high-risk regions where errors are likely to occur. Compared to traditional decoders that do not utilize temporal information (such as MWPM), this embodiment reduces the probability of missed and false detections, thereby reducing the logical error rate and improving quantum error correction accuracy. By pre-constructing a constrained decoding space, the decoder's search range is significantly reduced from the entire space, resulting in a significant decrease in the time consumption of error chain matching and solving, thus improving quantum error correction efficiency. Simultaneously, the backoff mechanism ensures that decoding can still be completed in the worst-case scenario (prediction failure), preventing indefinite stagnation due to prediction errors and enhancing the robustness of the quantum error correction system.

[0077] Meanwhile, the aforementioned quantum error correction device 300 is configured to execute any of the aforementioned quantum error correction methods. Therefore, the relevant modules in the aforementioned quantum error correction device are also used to execute the corresponding operations in any of the aforementioned quantum error correction methods. Accordingly, it also possesses all the beneficial effects of any of the aforementioned quantum error correction methods, which will not be elaborated upon here.

[0078] It should be noted that the specific modules within the aforementioned quantum error correction device are defined primarily based on the corresponding operations performed, and are not intended as limitations on the specific modules.

[0079] Furthermore, embodiments of the present invention also provide a quantum computing error correction system, comprising: At least one quantum processor is used to execute a surface code error correction loop and generate a discrete syndrome measurement sequence; At least one of the aforementioned quantum error correction devices, wherein the quantum error correction device is coupled to a quantum processor; The quantum error correction device includes a field-programmable gate array (FPGA), and the causal convolutional neural network is deployed in the logic resources of the FPGA.

[0080] The quantum computing error correction system includes the aforementioned quantum error correction device, and therefore has all the beneficial effects of the aforementioned quantum error correction device, which will not be elaborated further in this application.

[0081] This invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of this invention.

[0082] This invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the methods of embodiments of this invention. The computer program product should be understood as a software product that primarily implements the methods of this invention through a computer program.

[0083] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of this invention.

[0084] refer to Figure 4 The present invention will now be described in the form of a structural block diagram of an electronic device that can serve as an embodiment of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0085] like Figure 4 As shown, the electronic device includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0086] Multiple components in the electronic device are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information into the electronic device. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0087] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as a computer program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0088] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of embodiments of the present invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0090] It should be noted that the term "comprising" and its variations used in the embodiments of the present invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "multiple" mentioned in the embodiments of the present invention are illustrative and not restrictive. Those skilled in the art should understand that, unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0091] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in the embodiments of the present invention are all information and data that have been permitted by the user or have been fully agreed upon by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to agree or refuse.

[0092] The steps described in the method embodiments provided by this invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of this invention is not limited in this respect.

[0093] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A quantum error correction method, characterized by, include: Acquire discrete syndrome measurement sequences generated by a quantum processor based on surface codes over multiple consecutive error correction cycles; The discrete syndrome measurement sequence is input into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of qubit errors. Based on the probability distribution, a set of high-risk qubits is determined for the next error correction cycle, and a restricted decoding space for the next error correction cycle is pre-constructed based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; In the next error correction cycle, the control decoder prioritizes error chain matching within the restricted decoding space; if no error chain satisfying the verification conditions is found within the restricted decoding space, the control decoder backtracks to the full search space for matching; and outputs error correction operation instructions based on the matching results.

2. The method according to claim 1, characterized in that, The causal convolutional neural network comprises multiple levels of one-dimensional convolutional layers, and the receptive range of the convolutional kernels only covers the current and past time steps; wherein, The causal convolutional neural network is deployed in a field-programmable gate array to perform real-time inference on streaming discrete syndrome measurement sequences.

3. The method according to claim 2, characterized in that, The causal convolutional neural network contains multiple levels of dilated convolutional layers, with the dilation coefficient of each level increasing sequentially to expand the receptive field of the causal convolutional neural network to cover long-term noise correlation. The network structure of the causal convolutional neural network achieves temporal causality through the unidirectional coverage characteristic of the convolutional kernel's time axis. During the training phase of the causal convolutional neural network, the training samples consist of the historical syndrome window and the true error location labels of its subsequent error correction cycles. The network structure also restricts any predicted output to depend only on the current and historical syndrome inputs.

4. The method according to claim 1, characterized in that, The step of determining the set of high-risk qubits in the next error correction cycle based on the probability distribution includes: Calculate the prediction confidence level of the probability distribution; If the predicted confidence level is greater than a preset confidence threshold, then the qubits whose probability values ​​exceed the preset probability threshold are marked as high-risk qubits; wherein, the preset confidence threshold is dynamically adjusted according to the decoding success rate of historical error correction cycles.

5. The method according to claim 4, characterized in that, Also includes: If the prediction confidence is less than or equal to the preset confidence threshold, the construction step of the restricted decoding space is skipped, and the decoder is directly controlled to solve the error chain matching problem in the full search space in the next error correction cycle.

6. The method according to claim 1, characterized in that, Before inputting the discrete syndrome measurement sequence into the causal convolutional neural network, the method further includes: Obtain the physical location coordinates of each qubit on the quantum chip; The physical location coordinate information is fused with the discrete syndrome measurement sequence and used as a joint input to a causal convolutional neural network, so that the causal convolutional neural network can perceive non-uniform spatial noise distribution.

7. The method according to claim 1, characterized in that, The training steps of the causal convolutional neural network include: A training dataset is generated using a quantum error correction simulator; during the generation of the training dataset, random perturbations are applied to the physical bit error rate parameter to simulate quantum error correction environments under different noise intensities; Based on the physical bit error rate parameter after applying random perturbation, sample pairs containing discrete syndrome measurement sequences and corresponding true error locations are generated; The sample pairs are input into the causal convolutional neural network to be trained, and the difference between the predicted error probability distribution and the true error location is used as the loss function to update the parameters of the causal convolutional neural network.

8. A quantum error correction device, characterized in that, include: The acquisition module is used to acquire discrete syndrome measurement sequences generated by the quantum processor based on surface codes over multiple consecutive error correction cycles; The prediction module is used to input the discrete syndrome measurement sequence into a causal convolutional neural network to output error prediction information for the next error correction cycle; wherein, the error prediction information includes the probability distribution of qubit errors. A construction module is used to determine the set of high-risk qubits in the next error correction cycle based on the probability distribution, and to pre-construct a restricted decoding space for the next error correction cycle based on the set of high-risk qubits; wherein the restricted decoding space is a subset of the full search space; The error correction module is used to control the decoder to prioritize error chain matching and solving within the restricted decoding space in the next error correction cycle; if no error chain satisfying the verification condition is matched within the restricted decoding space, the decoder is controlled to fall back to the full search space for matching and solving; and the error correction operation instruction is output according to the matching and solving result.

9. A quantum computing error correction system, characterized in that, include: At least one quantum processor is used to execute a surface code error correction loop and generate a discrete syndrome measurement sequence; At least one quantum error correction device as described in claim 8, wherein the quantum error correction device is coupled to the quantum processor; The quantum error correction device is equipped with a field-programmable gate array (FPGA), and the causal convolutional neural network is deployed in the logic resources of the FPGA.

10. A non-transitory machine-readable medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 7.

Images