Four-point ramsey qubit parameter initialization method and system

Abstract

This invention relates to the fields of signal processing and control optimization, specifically to a method and system for initializing four-point Ramsey qubit parameters. The method receives measurement and control parameter boundaries, hardware constraints, readout assignment matrix versions, template versions, and hardware state summaries. Candidate parameters are generated through Latin hypercube coarse screening. Under the same hardware state and readout model, a dual-delay, biorthogonal four-point Ramsey template is executed to remove, discriminate, and correct anomalies in the IQ data, calculating the Ramsey fidelity and optimization objective. Interruption recovery is achieved through Bayesian fine-tuning and optimized state checkpoints, and the optimal initialization parameters are written back. This scheme improves parameter initialization efficiency, reduces the impact of readout errors and hardware drift on optimization, and enhances the reliability of subsequent runs after interruption.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and control optimization, specifically to a method and system for initializing four-point Ramsey qubit parameters. Background Technology

[0002] In superconducting quantum computing systems, qubit state control typically relies on an arbitrary waveform generator, a digital-to-analog converter, a microwave source, an IQ modulator, and related pulse compilation links to generate microwave pulse sequences for qubit control. Qubit state readout typically relies on a readout resonant cavity, an analog-to-digital converter, a data acquisition card, and an IQ discrimination link to sample and discriminate the returned readout waveform. To ensure that the qubit can stably execute subsequent quantum gate control, readout, or experimental batch processing tasks, the driving frequency, driving amplitude, and other parameters need to be determined before the experiment or during automatic calibration. Pulse amplitude, DRAG correction ratio, readout frequency, readout amplitude, readout duration, and parameters related to Ramsey measurements such as wait delay, phase, and readout window.

[0003] Existing manual parameter tuning or grid scanning methods typically require searching across multiple parameter dimensions with a fixed step size. When the dimension of the parameters to be initialized is high, the number of experiments increases rapidly with the parameter dimension and resolution, resulting in long calibration times, excessive manual intervention, and a tendency to miss effective parameter combinations in unexplored regions. While random search methods can reduce some of the overhead of fixed grids, their convergence stability is significantly affected by sampling density, step size, and initial range.

[0004] Bayesian optimization can be used for black-box parameter optimization, but in superconducting qubit measurement and control scenarios, each candidate parameter evaluation requires the generation of pulses, qubit responses, ADC sampling, and IQ discrimination through actual hardware, resulting in high experimental costs and significant observational noise. If Bayesian optimization lacks available initial observation points, it is prone to problems such as low exploration efficiency or local convergence when early samples are insufficient. On the other hand, qubit measurement and control hardware may experience interruptions, drift, or readout model changes during long-term operation. If only ordinary logs or historical results are saved, without saving the optimization state, template version, readout model version, and hardware state summary, invalid surrogate models or historical observation data may be incorrectly reused after the interruption is recovered.

[0005] The Ramsey experiment can characterize physical quantities such as qubit frequency shift, phase accumulation, and visibility. However, if only a few Ramsey delays or phase points are loosely combined into a scoring function, problems such as inconsistent evaluation structures between different candidate points, physical degradation of phase points, uncorrected readout errors, and unstable scoring results can easily arise. This is especially true when the phase setting only manifests as... When the command phase is periodically equivalent, the newly added measurement points may not provide independent physical observations, thus affecting the reliability of the optimization target. Summary of the Invention

[0006] The purpose of this invention is to provide a solution to the aforementioned problems existing in the prior art. Specifically, this invention is achieved through the following technical solution: The four-point Ramsey qubit parameter initialization method includes the following steps: Step 1: Receive the parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization target formula version, and current hardware state summary of the quantum bit measurement and control parameters, and construct candidate parameter mapping rules. Step 2: Perform Latin hypercube coarse screening sampling on the key dimensions to obtain multiple candidate parameters, and convert the multiple candidate parameters into executable pulse configurations and readout configurations respectively; Step 3: For the current candidate parameter among the multiple candidate parameters, under the same hardware state, the same readout assignment matrix version and the same IQ discriminator version represented by the current hardware state summary, execute a unified four-point Ramsey experimental template consisting of four Ramsey points. The four Ramsey points include short-delay point pairs and long-delay point pairs, and there is an orthogonal projection relationship between the two point pairs. Step 4: Perform outlier removal, IQ discrimination, and readout assignment matrix correction on the IQ sampling data of the four Ramsey points to obtain four corrected excited state probabilities corresponding to the four Ramsey points. Based on the four corrected excited state probabilities and the current candidate parameters, calculate short-delay visibility, long-delay visibility, Ramsey fidelity, and optimization objectives including normalized guidance penalty. Step 5: Construct a Bayesian fine-tuned surrogate model using the historical candidate parameters and historical optimization objectives from the coarse screening stage. Iterate through the expected improvement-anchor joint acquisition function to generate candidate parameters and update the surrogate model until the stopping condition is met. Step Six: During the coarse screening and fine-tuning process, generate optimization status checkpoints containing historical candidate parameters, historical optimization objectives, current optimal parameters, surrogate model status, candidate queue, random seed, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization objective formula version, readout assignment matrix version, IQ discriminator version, and current hardware status summary. During interrupt recovery, first perform integrity verification, unified four-point Ramsey experiment template version verification, Ramsey fidelity and optimization objective formula version verification, readout assignment matrix version verification, IQ discriminator version verification, hardware status summary comparison, and sentinel point retesting. Before the verification is completed, do not directly reuse the surrogate model status, and then perform full recovery, half recovery, or partial restart according to the verification results. Step 7: Write the final optimal parameters back to the parameter register as the initialization parameter set.

[0007] Furthermore, the four Ramsey points are respectively , , and ,in: ; ; ; ; in, For short waiting delay, For long waiting delay, For the second The reference relative amplitude ratio of the pulse. For reference, DRAG correction ratio, To read the reference configuration, To read window parameters, and These are the sampling times for the short-delay point pair and the long-delay point pair, respectively. and It is a positive integer; , , and This indicates the target's total readout projection phase.

[0008] Furthermore, the four Ramsey points share the same AWG / DAC output channel, the same microwave drive channel, the same trigger clock, the same readout window family, and the same IQ discriminator version; the target total readout projection phase is determined by the Ramsey free evolution phase and the end... The pulse projection phase is determined jointly to avoid the situation where only the pulse projection phase is determined. Periodic equivalent command phase formation degenerates observation.

[0009] Furthermore, the outlier removal, IQ discrimination, and readout assignment matrix correction of the IQ sampling data of the four Ramsey points include: For each Ramsey point, abnormal sampling data points or abnormal sampling batches are identified based on statistical distance or batch deviation of the IQ sampling data, and the abnormal sampling data points or abnormal sampling batches are removed from the corresponding IQ sampling data. The IQ discriminator is used to perform state discrimination on the IQ sampling data after anomaly removal to obtain the original state probability vector of each Ramsey point; The original state probability vector is corrected using the readout assignment matrix to obtain the corrected state probability vector for each Ramsey point. The excited state probability in each corrected state probability vector is then used as the corrected excited state probability for the corresponding Ramsey point to obtain four corrected excited state probabilities corresponding to the four Ramsey points.

[0010] Furthermore, the calculation of short-delay visibility, long-delay visibility, and Ramsey fidelity based on the four corrected excited-state probabilities and the current candidate parameters includes: ; ; ; ; in, For the short-delay visibility, For the long-delay visibility, This is Ramsey's evaluation score before cropping. To optimize the Ramsey fidelity used in the target computation, Indicates will Crop to interval, , , , These are the corrected excited-state probabilities for the four Ramsey points, respectively. The readout contrast is determined based on the contrast of the corrected reference state. It is a regularization constant and , For short delay weights, For long-delay weights, the and stated It has positive weights, and .

[0011] Furthermore, when normalized bootstrapping penalty is enabled and the parameter dimension index set is... When not empty, the optimization objective is denoted as And construct it according to the following formula: ; ; ; in, To normalize the guidance of punishment, The penalty coefficient and , For a smaller, better set of parameter dimension indexes, the The corresponding set of physical parameters is , For the parameter dimension index set The parameter dimension index in the text. For the first Non-negative penalty weights in each parameter dimension, and , For the first The penalty index for each parameter dimension, For candidate parameters in the th Physical parameter values ​​in each parameter dimension For the first The lower bound of each parameter dimension. For the first The upper bound of each parameter dimension. For the first The normalized parameter values ​​of each parameter dimension; the parameter dimensions participating in the normalized guided penalty calculation satisfy... The fixed parameters are not involved in the normalized guidance penalty calculation.

[0012] Furthermore, the step of iteratively generating candidate parameters through the desired improvement-anchor joint acquisition function includes: Calculate candidate points based on the surrogate model Expected improvements And calculate the candidate point score according to the following formula. : ; Select candidate points that satisfy the parameter boundaries and score them. The largest candidate point is used as the next candidate parameter; where, For the exploration parameters of the function to be improved and , Represents the square of the Euclidean norm. The anchor point distance penalty coefficient and , It is a diagonal matrix, the The diagonal elements represent the search interval lengths for each parameter dimension involved in the acquisition function calculation, and all search interval lengths are greater than 0. These are the anchor point parameters confirmed during the coarse screening or recovery stage; after generating the next candidate parameter, hardware boundary trimming, discrete parameter rounding, and invalid point filtering are performed on the next candidate parameter.

[0013] Furthermore, the optimized state checkpoint also includes a checkpoint structure version, parameter boundary version, fixed parameter version, candidate parameter mapping rule version, hardware channel mapping, sampling rate, trigger timing, read window parameters, read error rate, read length, decoherence time summary, surrogate model hyperparameters, integrity check value, and recovery strategy flag; the optimized state checkpoint forms a recoverable state through temporary state writing, integrity verification, and atomic replacement or equivalent transaction commit.

[0014] Furthermore, the step of performing a full recovery, partial recovery, or partial restart based on the verification result includes: When the integrity check passes, the unified four-point Ramsey experimental template version is consistent, the Ramsey fidelity and optimization target formula version is consistent, the parameter boundary version is consistent, the readout assignment matrix version is consistent, the IQ discriminator version is consistent, the difference obtained from the hardware state summary comparison does not exceed the first hardware drift threshold, and the target drift obtained from the sentinel point retest does not exceed the first target drift threshold, the proxy model state and the candidate queue are loaded to perform the full recovery; When the integrity check passes, the historical candidate parameters are reusable, the partial restart condition is not met, and one of the following conditions is met, the historical candidate parameters are retained, and the proxy model is retrained based on the retested or refreshed target value to perform the semi-recovery: the readout assignment matrix version or the IQ discriminator version is inconsistent but can be restored to consistency by remeasurement or refresh of the readout model; the difference obtained by hardware state summary comparison exceeds the first hardware drift threshold but does not reach the second hardware drift threshold; the target drift obtained by the retest of the sentinel point exceeds the first target drift threshold but does not reach the second target drift threshold; The proxy model state is not reused and local candidate points are regenerated to perform the local restart when one of the following conditions is met: the integrity check fails; the unified four-point Ramsey experimental template version is inconsistent; the Ramsey fidelity and optimization target formula versions are inconsistent; the parameter boundary versions are inconsistent; the readout assignment matrix version or the IQ discriminator version is inconsistent and cannot be restored to consistency by re-measurement or refreshing the readout model; the difference obtained from the hardware state summary comparison reaches the second hardware drift threshold; the target drift obtained from the sentinel point remeasurement reaches the second target drift threshold. Wherein, the second hardware drift threshold is higher than the first hardware drift threshold, and the second target drift threshold is higher than the first target drift threshold; the sentinel point retest is performed based on the current unified four-point Ramsey experimental template version and the current Ramsey fidelity and optimization target formula version; if the Ramsey fidelity and optimization target formula versions are consistent, and the target drift obtained by the current optimal parameter or the anchor parameter point in the sentinel point retest does not reach the second target drift threshold, and the current optimal parameter or the anchor parameter point satisfies the current parameter boundary and hardware constraints, then local candidate points are regenerated around the current optimal parameter or the anchor parameter point that has passed the retest and satisfies the constraints; if the Ramsey fidelity and optimization target formula versions are inconsistent, but the current optimal parameter or the anchor parameter point obtains a valid target value based on the current formula retest, is not marked as a low-confidence point for evaluation, and satisfies the current parameter boundary and hardware constraints, then local candidate points are regenerated around the parameter point; otherwise, local Latin hypercube sampling is performed according to the current parameter boundary to generate local candidate points.

[0015] The four-point Ramsey qubit parameter initialization system is used to execute the four-point Ramsey qubit parameter initialization method, including: a coarse screening sampling and parameter mapping module, a four-point Ramsey template execution module, an IQ discrimination and readout correction module, a Ramsey fidelity calculation module, a proxy optimization module, a checkpoint management module, a hardware consistency recovery module, a parameter register write-back module, and a data processing module. The coarse screening sampling and parameter mapping module, the four-point Ramsey template execution module, the IQ discrimination and readout correction module, the Ramsey fidelity calculation module, the proxy optimization module, the checkpoint management module, the hardware consistency recovery module, and the parameter register write-back module are respectively connected to the data processing module. The coarse screening and parameter mapping module is used to receive the parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experimental template version, Ramsey fidelity and optimization target formula version, and current hardware status summary of the quantum bit measurement and control parameters. It performs Latin hypercube coarse screening on the key dimensions and maps the candidate parameters to pulse configuration and readout configuration. The four-point Ramsey template execution module is used to execute a unified four-point Ramsey experimental template according to the pulse configuration and the readout configuration, under the same candidate parameters, the same hardware state represented by the current hardware state summary, the same readout assignment matrix version, and the same IQ discriminator version, and output IQ sampling data. The IQ discrimination and readout correction module is used to remove outliers, discriminate IQ data, and correct the readout assignment matrix to output four corrected excited state probabilities. The Ramsey fidelity calculation module is used to calculate short-delay visibility, long-delay visibility, and Ramsey fidelity based on the four corrected excited-state probabilities and corresponding candidate parameters. and optimization goals ; The proxy optimization module is used to construct a proxy model based on historical candidate parameters and historical optimization objectives, and to generate the next candidate parameters through the expected improvement-anchor point joint acquisition function; The checkpoint management module is used to generate optimized state checkpoints during the coarse screening and fine adjustment processes; The hardware consistency recovery module is used to perform integrity verification, unified four-point Ramsey experiment template version verification, Ramsey fidelity and optimization target formula version verification, read assignment matrix version verification, IQ discriminator version verification, hardware state summary comparison and sentinel point retesting before directly reusing the proxy model state when an interruption occurs and recovery is performed. Based on the verification results, the module controls the proxy optimization module to perform full recovery, half recovery or partial restart. The parameter register write-back module is used to write back the final optimal parameters as the initialization parameter set to the parameter register of the quantum bit measurement and control system.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: By employing a field-based, dual-delay, biorthogonal, unified four-point Ramsey experimental template, four Ramsey points under the same candidate parameters are executed under the same hardware state summary, the same trigger clock, and the same readout model version. This improves the comparability of scoring results between different candidate parameters and reduces the impact of loose phase points or... The risk of physical degradation caused by periodic equivalent phases.

[0017] The corrected excited state probability is obtained by IQ discrimination, outlier removal and readout assignment matrix correction, and Ramsey fidelity is obtained by aggregating short-delay visibility and long-delay visibility. This allows the coarse screening and fine-tuning stages to share the same type of physical evaluation quantity, reducing the impact of readout drift and outlier sampling on the optimization direction.

[0018] By using Latin hypercube coarse sampling to form a comprehensive initial observation point, and then using the expected improvement-anchor point joint acquisition function to sort the Bayesian fine-tuning candidate points, it is beneficial to reduce the risk of cold start in pure Bayesian optimization and increase the probability of entering the effective parameter region in the early stage of fine-tuning.

[0019] By incorporating smaller, better amplitude, duration, or perturbation parameters into the optimization objective through normalization-guided penalties, it is beneficial to alleviate the bias problem when optimizing parameters of different dimensions together.

[0020] By using optimization state checkpoints that include optimization state, template version, readout model version, and hardware state summary, and by performing sentinel point retests and hardware consistency recovery gates before recovery, it is beneficial to reduce repeated experiments after an interruption and avoid erroneous reuse of historical proxy models after significant drift in hardware state or readout model. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the four-point Ramsey qubit parameter initialization method. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are for illustrative purposes only and are not intended to limit the invention. It should be noted that this invention is already in the actual research and development stage.

[0023] Example 1 like Figure 1 As shown, the four-point Ramsey qubit parameter initialization method includes the following steps: Step 1: Receive the parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization target formula version, and current hardware state summary of the quantum bit measurement and control parameters, and construct candidate parameter mapping rules. Step 2: Perform Latin hypercube coarse screening sampling on the key dimensions to obtain multiple candidate parameters, and convert the multiple candidate parameters into executable pulse configurations and readout configurations respectively; Step 3: For the current candidate parameter among the multiple candidate parameters, under the same hardware state, the same readout assignment matrix version and the same IQ discriminator version represented by the current hardware state summary, execute a unified four-point Ramsey experimental template consisting of four Ramsey points. The four Ramsey points include short-delay point pairs and long-delay point pairs, and there is an orthogonal projection relationship between the two point pairs. Step 4: Perform outlier removal, IQ discrimination, and readout assignment matrix correction on the IQ sampling data of the four Ramsey points to obtain four corrected excited state probabilities corresponding to the four Ramsey points. Based on the four corrected excited state probabilities and the current candidate parameters, calculate short-delay visibility, long-delay visibility, Ramsey fidelity, and optimization objectives including normalized guidance penalty. Step 5: Construct a Bayesian fine-tuned surrogate model using the historical candidate parameters and historical optimization objectives from the coarse screening stage. Iterate through the expected improvement-anchor joint acquisition function to generate candidate parameters and update the surrogate model until the stopping condition is met. Step Six: During the coarse screening and fine-tuning process, generate optimization status checkpoints containing historical candidate parameters, historical optimization objectives, current optimal parameters, surrogate model status, candidate queue, random seed, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization objective formula version, readout assignment matrix version, IQ discriminator version, and current hardware status summary. During interrupt recovery, first perform integrity verification, unified four-point Ramsey experiment template version verification, Ramsey fidelity and optimization objective formula version verification, readout assignment matrix version verification, IQ discriminator version verification, hardware status summary comparison, and sentinel point retesting. Before the verification is completed, do not directly reuse the surrogate model status, and then perform full recovery, half recovery, or partial restart according to the verification results. Step 7: Write the final optimal parameters back to the parameter register as the initialization parameter set.

[0024] Specifically, in this embodiment, the quantum bit measurement and control parameters may include driving frequency, driving amplitude, and Pulse amplitude, second The parameters include pulse relative amplitude ratio, DRAG correction ratio, readout frequency, readout amplitude, readout duration, readout window, Ramsey wait delay, readout projection phase, and trigger parameters related to the hardware channel. Parameter boundaries characterize the upper and lower bounds, units, safety range, and hardware-executable range of each parameter to be optimized; key dimensions characterize the parameter dimensions involved in Latin hypercube coarse screening and Bayesian fine-tuning; fixed parameters characterize parameters that remain unchanged in the current initialization task but still need to be written into the experimental configuration; hardware constraints include hardware resolution and safety boundaries, where hardware resolution may include frequency step, amplitude quantization step, duration resolution, phase resolution, and rounding rules for discrete integer parameters.

[0025] The quantum bit measurement and control system receives parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization target formula version, and current hardware state summary. The current hardware state summary characterizes the measurement and control hardware state at the corresponding evaluation time and may include at least one of the following: backend identifier, calibration timestamp, AWG / DAC channel mapping, ADC sampling configuration, trigger timing, sampling rate, readout window parameters, readout error rate, readout length, decoherence time summary, microwave source frequency configuration, firmware version, and related configuration hash values. The current hardware state summary and candidate parameter evaluation results are jointly written into the optimization state checkpoint, used to determine whether historical observation data and surrogate model states can still be reused during recovery.

[0026] The readout model includes at least one of the following: a readout assignment matrix, an IQ discriminator, a readout reference configuration, and readout window parameters. The readout model version can be jointly determined by the readout assignment matrix version, the IQ discriminator version, the readout reference configuration version, and the readout window parameter version. The readout assignment matrix version is used to identify the current readout assignment matrix generation batch, the readout reference configuration, and the readout window parameters; the Ramsey fidelity and optimization target formula version is used to identify the current Ramsey fidelity calculation process, optimization target calculation process, normalization guidance penalty calculation process, and expected improvement-anchor joint acquisition function calculation process. The readout assignment matrix version and the Ramsey fidelity and optimization target formula version can be read from the initialization task configuration, historical calibration records, or the measurement and control system configuration file, and written into the optimization state checkpoint when generating the optimization state checkpoint.

[0027] In this embodiment, the system first performs Latin hypercube coarse screening sampling on the key dimensions to obtain multiple normalized candidate points. Let the set of key dimensions participating in the Latin hypercube coarse screening sampling be . For the first One key dimension Normalized parameters With physical parameters The following relationships can be used to map them together: ; in, For the first The physical parameter values ​​of the key dimensions, For the first Normalized parameter values ​​for each key dimension For the first The lower bound of each key dimension For the first The upper bound of each key dimension. The key dimensions participating in normalization satisfy... For discrete or integer parameters, the system rounds them according to hardware resolution or integer rounding rules after mapping. For candidate parameters that do not meet the safety boundaries, the system performs boundary pruning, resampling, or invalid point marking.

[0028] For each candidate parameter, the system writes the candidate parameter into the pulse configuration and readout configuration, and executes a unified four-point Ramsey experimental template under the same hardware state, the same readout assignment matrix version, and the same IQ discriminator version represented by the current hardware state summary. After the four-point Ramsey template is executed, the ADC or acquisition card outputs the IQ sampling data corresponding to each Ramsey point. The system performs outlier removal, IQ discrimination, and readout assignment matrix correction on the IQ sampling data to obtain the corrected excited-state probability. The system calculates short-delay visibility, long-delay visibility, and Ramsey fidelity based on the corrected excited-state probability and the current candidate parameters, and constructs an optimization objective based on the Ramsey fidelity and the current candidate parameters. The Ramsey fidelity in this application... The normalized evaluation value, constructed based on the corrected excited-state probabilities obtained from the unified four-point Ramsey experimental template, is used to measure the visibility performance of candidate parameters under the unified four-point Ramsey experimental template; the Ramsey fidelity is... It is not equivalent to quantum gate fidelity, readout fidelity, or quantum state fidelity.

[0029] After the initial screening stage, the system uses the historical candidate parameters and historical optimization objectives obtained from the initial screening as the initial observation data for Bayesian fine-tuning. The system constructs a surrogate model and generates the next candidate parameters through the expected improvement-anchor joint acquisition function. The surrogate model can be a random forest, an extreme random tree, a Gaussian process, or a gradient boosting regression tree. For scenarios with low dimensionality and most parameters being continuous, a Gaussian process can be used as the surrogate model; for scenarios with integer mappings, invalid points, and conditional parameters, a random forest or an extreme random tree can be used as the surrogate model.

[0030] During fine-tuning, the system repeatedly performs four-point Ramsey template processing, readout correction, Ramsey fidelity calculation, optimization target calculation, and surrogate model update for each new candidate parameter. Based on the updated surrogate model, the system again generates candidate parameters through the expected improvement-anchor joint acquisition function until a stopping condition is met. The stopping condition may include reaching a preset Ramsey fidelity threshold, reaching the maximum number of evaluations, failing to achieve the target improvement after a certain number of consecutive evaluations, user interruption, or hardware malfunction.

[0031] In one embodiment, the system can use the following parameter table to describe the parameters involved in the initialization task. The boundaries and resolutions of each parameter in the table are given by the configuration of the quantum bit measurement and control platform, historical calibration records, or experimental task configuration.

[0032] In the same initialization task, a certain parameter dimension is configured as either a critical dimension or a fixed parameter. When the parameter dimension is configured as a critical dimension, it participates in Latin hypercube coarse screening, Bayesian fine-tuning, and necessary normalization calculations. When the parameter dimension is configured as a fixed parameter, it does not participate in Latin hypercube coarse screening, expected improvement-anchor joint acquisition function distance calculation, or normalization-guided penalty calculation, but is still written into the corresponding pulse configuration or readout configuration according to the fixed parameter value.

[0033] In one embodiment, each Ramsey point in the unified four-point Ramsey experimental template is denoted as: ; in, For Ramsey point number, ; To wait for delay; For total readout projection phase; For the second Pulse relative amplitude ratio; This is the DRAG correction ratio; To read the reference configuration; To read out window parameters; The number of samples.

[0034] In a double-delay, double-orthogonal embodiment, the four Ramsey points can be configured as follows: ; ; ; ; in, and Forming short-delay point pairs, and Forming long-delay point pairs; For short waiting delay; For long waiting delays; For reference range ratio; For reference, DRAG correction ratio; To read the reference configuration; To read out window parameters; The number of samples for short-delay point pairs; This represents the number of samples for long-delay point pairs.

[0035] In the above embodiment, the four Ramsey points share the same readout reference configuration and the same readout window parameters, i.e. , The number of samples corresponding to short delay point pairs is... The number of samples corresponding to long delay point pairs is . and They are respectively and Implementation method when taking the same reference value; when the second When the pulse relative amplitude ratio or DRAG correction ratio is used as the parameter to be optimized... and Generated by candidate parameter mapping rules.

[0036] Short waiting delay Used to provide relatively stable comparative information on the readout link and low-latency phase response. Long wait delay This is used to provide observational information that is more sensitive to the effects of phase shift, frequency mismatch, or decoherence. The short waiting delay... and the aforementioned long waiting delay The parameters are determined based on the hardware clock resolution, decoherence time summary, and experimental task configuration.

[0037] and Having the same short waiting delay However, the target always reads out the projected phase difference. . and Having the same long waiting delay However, the target always reads out the projected phase difference. Long delay point pairs have, relative to short delay point pairs, The orthogonal projection relationship. With the above configuration, the four Ramsey points are not a simple stacking of any four Ramsey experiments, but rather a combination of short-delay contrast and long-delay phase sensitivity under the same candidate parameters, the same hardware state summary, the same readout assignment matrix version, and the same IQ discriminator version.

[0038] Four-point template , , and This indicates the target's total readout projection phase, not just the end phase. The command phase of the pulse. The total readout projection phase of the target is formed by the phase accumulation during Ramsey free evolution and the terminal phase. The pulse projection phase is jointly determined. The hardware command phase can be obtained by compensating the total target readout projection phase based on the estimated free-evolution phase. By using the total target readout projection phase, it is possible to avoid simply changing... Degenerate observations generated when the periodic equivalent command phase is used.

[0039] When executing the four-point Ramsey template, the pulse control system generates an initial pulse for each Ramsey point. Pulse, wait delay, terminal The pulse sequence of the pulse and the readout pulse. The second one... The pulse phase, amplitude ratio, and DRAG correction ratio are based on In , and Compile. Read window parameters. Used to control the ADC sampling window, integration window, or IQ demodulation window. The four Ramsey points share the same trigger clock and the same readout discrimination link to ensure the comparability of Ramsey fidelity between different candidate parameters.

[0040] In another embodiment, and The impact of candidate parameter changes on Ramsey visibility can be assessed by combining the amplitude ratio and DRAG correction. At this point, the four Ramsey points still maintain a double-delay, double-orthogonal structure, but the candidate parameter mapping rules will... and As a parameter to be optimized or a pulse control quantity derived from a parameter to be optimized.

[0041] Example 2 For each Ramsey point, the system collects data. There are 10 IQ samples. Each IQ sample can be represented as a two-dimensional vector. ,in, For in-phase components, For orthogonal components, This refers to the sampling data point number. Each sampling data point corresponds to an IQ sample obtained from one repeated measurement, and each sampling batch corresponds to a group of IQ samples divided according to a time window or sampling order.

[0042] The system first performs outlier removal on the IQ samples. Outlier removal can be based on statistical distance, Mahalanobis distance, absolute deviation of the median, or deviation of the batch mean. For sampled data points or batches that are determined to be outliers, the system removes them from the IQ sample set corresponding to the Ramsey point, or marks the Ramsey point as a low-confidence point for evaluation.

[0043] After outlier removal, the system uses an IQ discriminator to classify the state of the IQ samples. The IQ discriminator can be a threshold discriminator, a linear discriminator, a Gaussian mixture discriminator, or a classifier trained on calibration data. The IQ discriminator outputs the original state probability vector for each Ramsey point. : ; in, For the first The probability of a Ramsey point being classified as the ground state. For the first The probability of a Ramsey point being classified as an excited state, indicated by the superscript. This represents the transpose of a vector.

[0044] Read the assignment matrix The assignment relationship between the true state and the discrimination label can be represented as: ; in, Indicates the true state as When judged as The probability, and These represent the ground state label or the excited state label, respectively.

[0045] The readout assignment matrix This can be obtained under a readout reference configuration. Specifically, the qubit measurement and control system collects IQ samples under both ground-state and excited-state reference conditions, and the IQ discriminator outputs a discrimination tag; the readout assignment matrix is ​​constructed based on the statistical relationship between the actual reference state and the discrimination tag. The readout assignment matrix The version number, along with the IQ discriminator version, readout window parameters, and hardware status summary, is written into the optimized state checkpoint. During recovery, if the current readout assignment matrix version is inconsistent with the readout assignment matrix version saved in the checkpoint, or if the difference between the current readout assignment matrix and the readout assignment matrix saved in the checkpoint exceeds a preset readout model threshold, then a semi-recovery or partial restart determination process is initiated.

[0046] In this embodiment, the original state probability vector and the corrected state probability vector satisfy: ; Therefore, in When it is reversible, we get: ; in, For the first The corrected state probability vector of Ramsey points To read the assignment matrix The inverse matrix. When The number of conditions exceeds the preset condition number threshold or If the reversibility condition is not met, the system can use pseudo-inverse correction, remeasure and read out the reference configuration, or mark the candidate parameter as an invalid point for evaluation.

[0047] If the corrected probability component exceeds If the probability deviates from 1, the system performs non-negative pruning and normalization, or marks the evaluation of that point as a low confidence score.

[0048] No. The corrected excited-state probabilities of the Ramsey points are denoted as: ; In this context, the subscript 1 indicates the excited state label.

[0049] For a dual-delay, biorthogonal four-point template, when the short-delay visibility and long-delay visibility are calculated using the corrected excited-state probability, the readout contrast... The corrected reference-state contrast is used and can be obtained according to the following formula: ; in, The excited state probabilities are obtained after correction by the readout assignment matrix under the excited state reference conditions. This represents the excited state probability obtained after correction using the readout assignment matrix under the ground state reference condition. Using the corrected reference state contrast can... It maintains the same correction caliber as the corrected excited-state probability used to calculate visibility. When When the readout contrast falls below the preset threshold, the current readout model is marked as low confidence, triggering readout model refresh, reference state retest, or low confidence scoring processing.

[0050] The system calculates short-latency visibility using the following formula. and long-delay visibility : ; ; in, and These are the corrected excited-state probabilities for short-delay point pairs, and These are the corrected excited-state probabilities for long-delay point pairs, To prevent regularization constants with denominators of zero.

[0051] The system first calculates the Ramsey score before cropping using the following formula. : ; The system then obtains the Ramsey fidelity used to optimize the target calculation according to the following formula. : ; in, Indicates when When 0 is taken, Take 1 in the first case, and take 1 in the other cases. ; For short delay weights, For long-delay weights, and , , In one embodiment, the weights can be set as follows: ; ; in, The number of samples for short-delay point pairs. For the number of samples of long-delay point pairs, For long waiting delay, Decoherence time of qubits The estimated value, ,and When unavailable In this case, positive value estimation can be used from the platform task configuration, or preset positive weights can be used directly. Through the above weight settings, while providing phase sensitivity, the weight of long delay points is constrained by decoherence estimation to reduce the misleading effect on the optimization direction when the noise at long delay points is too large.

[0052] Unless otherwise stated, optimization objective Using trimmed and low-confidence processed When a Ramsey point corresponding to a candidate parameter is marked as a low-confidence point for evaluation, the system assigns a preset high target value to the candidate parameter by default, so that the surrogate model reduces the sampling priority of the neighborhood of the candidate parameter. If the initialization task configuration allows, the candidate parameter can also be resampled, or it can be removed from the surrogate model's training data. The low-confidence handling method adopted is written into the Ramsey fidelity and optimization target formula version. The preset high target value is determined by the initialization task configuration, historical invalid point evaluation statistics, or platform security policies; the preset high target value is not lower than the target value corresponding to invalid candidate points in the current optimization task, and is higher than the current optimal target value, so that the surrogate model reduces the sampling priority of the neighborhood of low-confidence candidate parameters.

[0053] Example 3 The system constructs optimization objectives by minimizing the problem. : ; in, To optimize the objective, To ensure the fidelity of Ramsey, To normalize the guiding penalty coefficient, and When normalized boot penalties are enabled, ; This is a normalization-guided penalty item.

[0054] When normalized bootstrapping penalty is enabled and the parameter dimension index set is smaller as much as possible. When not empty, the system calculates according to the following formula. : ; ; in, For a smaller, better set of parameter dimension indexes, the The corresponding set of physical parameters is ,and For the set of key dimensions A subset of, or a subset of the set of parameter dimension indexes involved in the optimization; For parameter dimension index, As a penalty weight, As a penalty index, For normalized parameter values, For candidate parameters in the th Physical parameter values ​​in each dimension For the first The lower bound of each dimension, For the first The upper bound of each dimension. The parameter dimensions involved in the normalization-guided penalty calculation satisfy... Fixed parameters do not participate in the normalized boot penalty calculation. This applies when normalized boot penalties are enabled and the parameter dimension index set... When not empty, the stated and at least one , making Preferably, The aforementioned The aforementioned Based on hardware boundary clipping and discretization... Calculation. When normalized boot penalty is not enabled, i.e. or the parameter dimension index set When it is an empty set, it is stipulated that .

[0055] Smaller is better parameter dimension index set The corresponding physical parameters may include drive amplitude offset, readout amplitude, readout duration, DRAG correction amplitude, or other parameters that need to be controlled to small values; the specific set is determined by the experimental platform and mission objectives. For physical parameters The smaller the value, the smaller the contribution of the normalization-guided penalty term to the optimization objective. (Through...) The objective function of the system is to maintain Ramsey fidelity. While maintaining the primary objective, normalization constraints are applied to small-scale optimization parameters with different dimensions to prevent unreasonable biases from arising during optimization due to the large dimension or wide numerical range of a particular parameter. The configuration should make Not exceeding the target shaping range configured in the initialization task, and with As the primary optimization objective.

[0056] In this application , , , , and This can be determined by platform calibration data, historical experiment statistics, or initialization task configuration. This is used to avoid a denominator of zero when the readout contrast is too low; the aforementioned , and Used to control the impact of a smaller, better set of parameter dimension indexes on the optimization objective; and These parameters are used to adjust the exploration level and anchor neighborhood constraint strength in the desired improvement of the joint acquisition function. Different qubit measurement and control platforms can adjust these parameters based on hardware security boundaries, parameter stability, and historical optimization results; parameter selection should not exceed hardware security boundaries and should not cause... The primary objective is suppressed by the penalty item as a constraint.

[0057] During the fine-tuning phase, the system uses historical candidate parameters and historical optimization objectives obtained in the coarse-screening phase as initial observation data to construct a surrogate model. Let the surrogate model be at the candidate points... The predicted mean given is The prediction standard deviation is The current optimal observation target value is The exploration parameters are ,and .when At that time, the system can calculate the expected improvement for minimizing the objective according to the following formula. : ; ; in, For standardized variables, The cumulative distribution function of the standard normal distribution. Let be the probability density function of the standard normal distribution. To avoid numerical instability caused by zero variance, when... hour, It can be set to 0, or the zero-variance handling rule in the initialization task configuration can be used. The above-mentioned desired improvement is based on minimizing... The direction is defined.

[0058] The system introduces an anchor point distance penalty and constructs an improved joint anchor point acquisition function: ; in, Score the candidate points. Represents the square of the Euclidean norm. The anchor point distance penalty coefficient and When anchor distance penalty is enabled, ; It is a diagonal matrix. The diagonal elements represent the search interval lengths for each parameter dimension involved in the acquisition function calculation, and all search interval lengths are greater than 0. These are anchor point parameter points. Anchor point parameter points can be the current optimal parameters confirmed during the coarse screening stage, or historically optimal parameters that remain reliable after retesting through sentinel points during the recovery stage. The system selects... The largest candidate point is selected as the next candidate parameter.

[0059] The and anchor point parameter points All parameters are encoded using physical parameters that conform to the candidate parameter mapping rules; fixed parameters are not included. The system constructs the discrete parameters; these parameters are first mapped according to hardware resolution before being used for scoring. After generating the next candidate parameter, the system performs boundary pruning, discrete parameter rounding, and invalid point filtering on the next candidate parameter. For candidate points that do not meet hardware security boundaries or readout configuration constraints, the system can resample, apply a high penalty target value, or add them to the invalid point set to avoid subsequent acquisition functions repeatedly proposing unexecutable candidate points.

[0060] During the coarse screening and fine-tuning processes, the system generates optimization state checkpoints according to preset evaluation intervals, preset time intervals, or key state update events. Each optimization state checkpoint includes at least the following: historical candidate parameters, historical optimization objectives, current optimal parameters, current optimal objective value, surrogate model state, candidate queue, random seed, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization objective formula version, readout assignment matrix version, IQ discriminator version, and a summary of the current hardware state.

[0061] In one embodiment, the optimized state checkpoint further includes a checkpoint structure version, parameter boundary version, fixed parameter version, candidate parameter mapping rule version, hardware channel mapping, sampling rate, trigger timing, readout window parameters, readout error rate, readout length, decoherence time summary, surrogate model hyperparameters, integrity check value, and recovery strategy flag.

[0062] Optimization status checkpoints can be saved using a combination of manifest files and historical data files. The manifest file stores the version, template, hardware status summary, and integrity verification fields; the historical data file stores historical candidate parameters, historical optimization goals, etc. , The system includes anomaly point ratio and validity markers; the proxy model state can be stored as a binary object or a reconfigurable parameter set, depending on the model type. When writing to optimized state checkpoints, the system can employ temporary state writing, storage refresh, integrity verification, and atomic replacement or equivalent transaction commit to reduce recovery failures caused by partial writes. The equivalent transaction commit refers to a write mechanism that guarantees the overall commit or rollback of fields in the optimized state checkpoint.

[0063] The Ramsey fidelity and optimization target formula version is used to identify... , , And the calculation rules for the expected improvement of the anchor point joint acquisition function. If any of the above calculation rules change, a corresponding version verification is required during recovery to determine whether the historical optimization objectives and the proxy model states are still comparable.

[0064] Upon interruption recovery, the system first loads the optimized state checkpoint and performs an integrity check before directly reusing the surrogate model state. After the integrity check passes, the system sequentially compares the checkpoint structure version, the unified four-point Ramsey experimental template version, the Ramsey fidelity and optimization objective formula version, the parameter boundary version, the readout assignment matrix version, and the IQ discriminator version. Subsequently, the system reads the current hardware state summary and compares it with the hardware state summary stored in the checkpoint. The hardware state summary comparison may include the backend identifier, calibration timestamp, channel mapping, sampling rate, trigger timing, readout window parameters, readout error rate, readout length, decoherence time summary, and related hash values.

[0065] The difference obtained from the hardware state summary comparison can be denoted as: For numerical fields, the system can normalize the field differences according to the corresponding threshold or historical standard deviation and then perform a weighted summation; for discrete fields, the system can convert the consistency comparison results into numerical trigger items; for fields with inconsistent hash configurations, the system can directly... The system will mark the second hardware drift threshold as reached. The recovery strategy is determined by comparing the results with a first hardware drift threshold and a second hardware drift threshold. A weighted summation method is used when calculating... At that time, the weights of each numerical field are determined by the initialization task configuration, historical calibration statistics, or platform maintenance strategy. Different platforms can adjust these weights based on hardware stability, readout drift frequency, and historical recovery results, but the weights and their versions should be written into the optimized state checkpoint or the corresponding recovery strategy flag. In one implementation, the difference score obtained from the hardware state summary comparison can be expressed as: ; in, This is a set of numerical hardware summary fields. For the first The difference between numerical fields after positive normalization scaling; when the historical standard deviation is 0 or unavailable, the positive scaling in the initialization task configuration is used as the normalization scaling; For the first The weights of each numeric field, and Preferred to satisfy ; Configure a discrete field or a value trigger for hash inconsistency. It can be 0 or a preset value not lower than the second hardware drift threshold; The preset value is determined by the recovery strategy flag, initialization task configuration, or historical recovery statistics; when inconsistent discrete fields or inconsistent configuration hashes are considered serious mismatches in the task configuration, You can directly take no less than The system determines whether to trigger a partial restart based on the specified value; when the configuration hash is inconsistent, the system can also directly determine the restart without calculating the weighted sum. .

[0066] After completing the version and hardware status summary comparison, the system re-executes the unified four-point Ramsey experiment template at one or more sentinel points to obtain the target value at the time of recovery. and compared with the corresponding historical target value in the checkpoint. By comparison, the target drift of the sentry point was obtained. : ; in, For sentry point target drift, The optimization objective obtained by retesting during recovery. The historical optimization targets are stored in the checkpoints. When multiple sentry points exist, the system calculates the targets for each sentry point separately. and its maximum value, median value, or weighted average value as The aggregation method used is determined by the recovery strategy flag and written into the optimization state checkpoint.

[0067] The recovery strategy can be determined hierarchically based on a first hardware drift threshold, a second hardware drift threshold, a first target drift threshold, and a second target drift threshold. The first hardware drift threshold is denoted as... The second hardware drift threshold is denoted as ,and The first target drift threshold is denoted as The second target drift threshold is denoted as ,and The system is based on and , The comparison results, and and , The comparison results determine the recovery strategy. The first hardware drift threshold and the first target drift threshold are used to determine whether the proxy model state and candidate queue can be directly reused; the second hardware drift threshold and the second target drift threshold are used to determine whether the current proxy model state must be abandoned and a partial restart must be performed.

[0068] When the integrity check passes, and the four Ramsey experimental template versions are consistent, the Ramsey fidelity and optimization target formula versions are consistent, the parameter boundary versions are consistent, the readout assignment matrix versions are consistent, and the IQ discriminator versions are consistent, then the integrity check is complete. ,and When the interruption occurs, the system performs a full recovery. During a full recovery, the system loads the agent model state, candidate queue, random seed, historical candidate parameters, and historical optimization objectives, and continues fine-tuning from the optimization state before the interruption.

[0069] When the integrity check passes, historical candidate parameters can be reused, the partial restart condition is not met, and the readout assignment matrix version or IQ discriminator version is inconsistent but can be restored to consistency by re-measuring or refreshing the readout model, or ,or During this process, the system performs a semi-recovery. In the semi-recovery process, the system prioritizes retaining historical candidate parameters; historical optimization targets are reused only if the formula version, readout model, and sentinel point verification pass, otherwise they are replaced with retested target values. The surrogate model is retrained based on the retested target values ​​and historical target values ​​that are still within the reusable range.

[0070] When integrity verification fails, or when the versions of the four unified Ramsey experimental templates are inconsistent, or when the versions of the Ramsey fidelity and optimization objective formulas are inconsistent, or when the versions of the parameter boundaries are inconsistent, or when the versions of the readout assignment matrix or the IQ discriminator are inconsistent and cannot be restored to consistency by re-measuring or refreshing the readout model. ,or When this happens, the system performs a partial restart. During a partial restart, the system does not reuse the proxy model state and regenerates local candidate points.

[0071] Sentinel point retesting uses the current unified four-point Ramsey experimental template version and the current Ramsey fidelity and optimization target formula version at the time of recovery. When the Ramsey fidelity and optimization target formula version saved in the checkpoint is consistent with the current version, if the target drift obtained by the current optimal parameter or anchor parameter point in the sentinel point retesting does not reach the second target drift threshold, and the current optimal parameter or anchor parameter point satisfies the current parameter boundary and hardware constraints, then the system regenerates local candidate points around the current optimal parameter or anchor parameter point that has passed the retesting and satisfies the constraints; when the Ramsey fidelity and optimization target formula version saved in the checkpoint is inconsistent with the current version, if the current optimal parameter or anchor parameter point obtains a valid target value based on the current formula retesting, is not marked as a low-confidence point for evaluation, and satisfies the current parameter boundary and hardware constraints, then the system regenerates local candidate points around that parameter point; otherwise, the system performs local Latin hypercube sampling based on the current parameter boundary to generate a set of local candidate points. Partial restarts do not force the reuse of expired proxy model states to avoid incorrect convergence after hardware state drift.

[0072] The readout model threshold is used to determine whether the readout model needs to be refreshed; the hardware drift threshold is used to determine the recovery level; and the target drift threshold is used to determine the sentinel point retest. If, during recovery, the readout assignment matrix version or the IQ discriminator version is found to be inconsistent with the version saved in the checkpoint, the system first performs a readout model refresh. The readout model refresh includes re-measuring the readout reference configuration, rebuilding the readout assignment matrix, or updating the IQ discriminator. If the refreshed readout model matches the current task configuration, and the target drift obtained from the sentinel point retest does not reach the second target drift threshold, then historical candidate parameters can be used as training input for semi-recovery; if the refreshed readout model cannot match the current task configuration, or the target drift obtained from the sentinel point retest reaches the second target drift threshold, then a partial restart is performed.

[0073] Example 4 This embodiment provides a four-point Ramsey qubit parameter initialization system. The system includes a coarse screening sampling and parameter mapping module, a four-point Ramsey template execution module, an IQ discrimination and readout correction module, a Ramsey fidelity calculation module, a proxy optimization module, a checkpoint management module, a hardware consistency recovery module, and a parameter register write-back module.

[0074] The coarse screening and parameter mapping module is used to receive parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization target formula version, and current hardware status summary. It performs Latin hypercube coarse screening on the key dimensions and maps candidate parameters to impulse configurations and readout configurations.

[0075] The four-point Ramsey template execution module is connected to the coarse screening sampling and parameter mapping module. The four-point Ramsey template execution module generates pulse sequences corresponding to the four Ramsey points based on the pulse configuration and readout configuration, and controls the AWG / DAC, microwave drive link, trigger clock, and readout acquisition link to execute a unified four-point Ramsey experimental template under the same candidate parameters, the same hardware state represented by the current hardware state summary, the same readout assignment matrix version, and the same IQ discriminator version. The four-point Ramsey template execution module outputs IQ sampling data.

[0076] The IQ discrimination and readout correction module is connected to the four-point Ramsey template execution module. The IQ discrimination and readout correction module performs outlier removal, IQ discrimination, and readout assignment matrix correction on the IQ sampled data to obtain four corrected excited state probabilities.

[0077] The Ramsey fidelity calculation module is connected to the IQ discrimination and readout correction module. The Ramsey fidelity calculation module calculates short-delay visibility, long-delay visibility, and Ramsey fidelity based on four corrected excited-state probabilities and corresponding candidate parameters. and optimization goals And will optimize the target Output to the proxy optimization module.

[0078] The proxy optimization module is connected to the Ramsey fidelity calculation module and the coarse screening sampling and parameter mapping module. The proxy optimization module constructs a proxy model based on historical candidate parameters and historical optimization objectives, generates the next candidate parameters through the expected improvement-anchor joint acquisition function, and sends the next candidate parameters to the coarse screening sampling and parameter mapping module for hardware mapping.

[0079] The checkpoint management module is connected to the proxy optimization module. During the coarse screening and fine-tuning processes, the checkpoint management module generates optimization status checkpoints and saves historical candidate parameters, historical optimization objectives, current optimal parameters, current optimal objective values, proxy model status, candidate queue, random seed, template version, formula version, readout model version, and a summary of the current hardware status.

[0080] The hardware consistency recovery module is connected to the checkpoint management module and the proxy optimization module. When an interruption occurs and recovery occurs, the hardware consistency recovery module reads the optimized state checkpoint and the current hardware state summary. Before directly reusing the proxy model state, it performs integrity checks, template version checks, Ramsey fidelity and optimization target formula version checks, read assignment matrix version checks, IQ discriminator version checks, hardware state summary comparisons, and sentinel point retests. Based on the verification results, the hardware consistency recovery module controls the proxy optimization module to perform full recovery, partial recovery, or partial restart.

[0081] The parameter register write-back module is connected to the proxy optimization module. The parameter register write-back module writes the final optimal parameters as the initialization parameter set back to the parameter register of the quantum bit measurement and control system. The parameter register may include a drive frequency register, a drive amplitude register, and... Pulse amplitude register, DRAG correction register, readout frequency register, readout amplitude register, readout duration register, and readout window configuration register.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for initializing Ramsey qubit parameters based on four points, characterized in that, Includes the following steps: Step 1: Receive the parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization target formula version, and current hardware state summary of the quantum bit measurement and control parameters, and construct candidate parameter mapping rules. Step 2: Perform Latin hypercube coarse screening sampling on the key dimensions to obtain multiple candidate parameters, and convert the multiple candidate parameters into executable pulse configurations and readout configurations respectively; Step 3: For the current candidate parameter among the multiple candidate parameters, under the same hardware state, the same readout assignment matrix version and the same IQ discriminator version represented by the current hardware state summary, execute a unified four-point Ramsey experimental template consisting of four Ramsey points. The four Ramsey points include short-delay point pairs and long-delay point pairs, and there is an orthogonal projection relationship between the two point pairs. Step 4: Perform outlier removal, IQ discrimination, and readout assignment matrix correction on the IQ sampling data of the four Ramsey points to obtain four corrected excited state probabilities corresponding to the four Ramsey points. Based on the four corrected excited state probabilities and the current candidate parameters, calculate short-delay visibility, long-delay visibility, Ramsey fidelity, and optimization objectives including normalized guidance penalty. Step 5: Construct a Bayesian fine-tuned surrogate model using the historical candidate parameters and historical optimization objectives from the coarse screening stage. Iterate through the expected improvement-anchor joint acquisition function to generate candidate parameters and update the surrogate model until the stopping condition is met. Step Six: During the coarse screening and fine-tuning process, generate optimization status checkpoints containing historical candidate parameters, historical optimization objectives, current optimal parameters, surrogate model status, candidate queue, random seed, unified four-point Ramsey experiment template version, Ramsey fidelity and optimization objective formula version, readout assignment matrix version, IQ discriminator version, and current hardware status summary. During interrupt recovery, first perform integrity verification, unified four-point Ramsey experiment template version verification, Ramsey fidelity and optimization objective formula version verification, readout assignment matrix version verification, IQ discriminator version verification, hardware status summary comparison, and sentinel point retesting. Before the verification is completed, do not directly reuse the surrogate model status, and then perform full recovery, half recovery, or partial restart according to the verification results. Step 7: Write the final optimal parameters back to the parameter register as the initialization parameter set.

2. The method for initializing the parameters of a four-point Ramsey qubit according to claim 1, characterized in that, The four Ramsey points are respectively , , and ,in: ； ； ； ； in, For short waiting delay, For long waiting delay, For the second The reference relative amplitude ratio of the pulse. For reference, DRAG correction ratio, To read the reference configuration, To read window parameters, and These are the sampling times for the short-delay point pair and the long-delay point pair, respectively. and It is a positive integer; , , and This indicates the target's total readout projection phase.

3. The method for initializing the parameters of a four-point Ramsey qubit according to claim 2, characterized in that, The four Ramsey points share the same AWG / DAC output channel, the same microwave drive channel, the same trigger clock, the same readout window family, and the same IQ discriminator version; the target total readout projection phase is determined by the Ramsey free evolution phase and the end... The pulse projection phase is determined jointly to avoid the situation where only the pulse projection phase is determined. Periodic equivalent command phase formation degenerates observation.

4. The method for initializing the parameters of a four-point Ramsey qubit according to claim 3, characterized in that, The outlier removal, IQ discrimination, and readout assignment matrix correction of the IQ sampling data of the four Ramsey points include: For each Ramsey point, abnormal sampling data points or abnormal sampling batches are identified based on statistical distance or batch deviation of the IQ sampling data, and the abnormal sampling data points or abnormal sampling batches are removed from the corresponding IQ sampling data. The IQ discriminator is used to perform state discrimination on the IQ sampling data after anomaly removal to obtain the original state probability vector of each Ramsey point; The original state probability vector is corrected using the readout assignment matrix to obtain the corrected state probability vector for each Ramsey point. The excited state probability in each corrected state probability vector is then used as the corrected excited state probability for the corresponding Ramsey point to obtain four corrected excited state probabilities corresponding to the four Ramsey points.

5. The method for initializing the parameters of a four-point Ramsey qubit according to claim 4, characterized in that, The calculation of short-delay visibility, long-delay visibility, and Ramsey fidelity based on the four corrected excited-state probabilities and the current candidate parameters includes: ； ； ； ； in, For the short-delay visibility, For the long-delay visibility, This is Ramsey's evaluation score before cropping. To optimize the Ramsey fidelity used in the target computation, Indicates will Crop to interval, , , , These are the corrected excited-state probabilities for the four Ramsey points, respectively. The readout contrast is determined based on the contrast of the corrected reference state. It is a regularization constant and , For short delay weights, For long-delay weights, the and stated It has positive weights, and .

6. The method for initializing the parameters of a four-point Ramsey qubit according to claim 5, characterized in that, When normalized bootstrapping penalty is enabled and the parameter dimension index set is... When not empty, the optimization objective is denoted as And construct it according to the following formula: ； ； ； in, To normalize the guidance of punishment, The penalty coefficient and , For a smaller, better set of parameter dimension indexes, the The corresponding set of physical parameters is , For the parameter dimension index set The parameter dimension index in the text. For the first Non-negative penalty weights in each parameter dimension, and , For the first The penalty index for each parameter dimension, For candidate parameters in the th Physical parameter values ​​in each parameter dimension For the first The lower bound of each parameter dimension. For the first The upper bound of each parameter dimension. For the first The normalized parameter values ​​of each parameter dimension; the parameter dimensions participating in the normalized guided penalty calculation satisfy... The fixed parameters are not involved in the normalized guidance penalty calculation.

7. The method for initializing the parameters of a four-point Ramsey qubit according to claim 6, characterized in that, The process of iteratively generating candidate parameters through the desired improvement-anchor joint acquisition function includes: Calculate candidate points based on the surrogate model Expected improvements And calculate the candidate point score according to the following formula. : ； Select candidate points that satisfy the parameter boundaries and score them. The largest candidate point is used as the next candidate parameter; where, For the exploration parameters of the function to be improved and , Represents the square of the Euclidean norm. The anchor point distance penalty coefficient and , It is a diagonal matrix, the The diagonal elements represent the search interval lengths for each parameter dimension involved in the acquisition function calculation, and all search interval lengths are greater than 0. Anchor point parameter points confirmed in the coarse screening or recovery stage; After generating the next candidate parameter, perform hardware boundary trimming, discrete parameter rounding, and invalid point filtering on the next candidate parameter.

8. The method for initializing the parameters of a four-point Ramsey qubit according to claim 7, characterized in that, The optimized state checkpoint also includes a checkpoint structure version, parameter boundary version, fixed parameter version, candidate parameter mapping rule version, hardware channel mapping, sampling rate, trigger timing, readout window parameters, readout error rate, readout length, decoherence time summary, surrogate model hyperparameters, integrity check value, and recovery strategy flag; the optimized state checkpoint forms a recoverable state through temporary state writing, integrity check, and atomic replacement or equivalent transaction commit.

9. The method for initializing the parameters of a four-point Ramsey qubit according to claim 8, characterized in that, The step of performing a full recovery, partial recovery, or partial restart based on the verification results includes: When the integrity check passes, the unified four-point Ramsey experimental template version is consistent, the Ramsey fidelity and optimization target formula version is consistent, the parameter boundary version is consistent, the readout assignment matrix version is consistent, the IQ discriminator version is consistent, the difference obtained from the hardware state summary comparison does not exceed the first hardware drift threshold, and the target drift obtained from the sentinel point retest does not exceed the first target drift threshold, the proxy model state and the candidate queue are loaded to perform the full recovery; When the integrity check passes, the historical candidate parameters are reusable, the partial restart condition is not met, and one of the following conditions is met, the historical candidate parameters are retained, and the proxy model is retrained based on the retested or refreshed target value to perform the semi-recovery: the readout assignment matrix version or the IQ discriminator version is inconsistent but can be restored to consistency by remeasurement or refresh of the readout model; the difference obtained by hardware state summary comparison exceeds the first hardware drift threshold but does not reach the second hardware drift threshold; the target drift obtained by the retest of the sentinel point exceeds the first target drift threshold but does not reach the second target drift threshold; The proxy model state is not reused and local candidate points are regenerated to perform the local restart when one of the following conditions is met: the integrity check fails; the unified four-point Ramsey experimental template version is inconsistent; the Ramsey fidelity and optimization target formula versions are inconsistent; the parameter boundary versions are inconsistent; the readout assignment matrix version or the IQ discriminator version is inconsistent and cannot be restored to consistency by re-measurement or refreshing the readout model; the difference obtained from the hardware state summary comparison reaches the second hardware drift threshold; the target drift obtained from the sentinel point remeasurement reaches the second target drift threshold. Wherein, the second hardware drift threshold is higher than the first hardware drift threshold, and the second target drift threshold is higher than the first target drift threshold; the sentinel point retest is performed based on the current unified four-point Ramsey experimental template version and the current Ramsey fidelity and optimization target formula version; if the Ramsey fidelity and optimization target formula versions are consistent, and the target drift obtained by the current optimal parameter or the anchor parameter point in the sentinel point retest does not reach the second target drift threshold, and the current optimal parameter or the anchor parameter point satisfies the current parameter boundary and hardware constraints, then local candidate points are regenerated around the current optimal parameter or the anchor parameter point that has passed the retest and satisfies the constraints; if the Ramsey fidelity and optimization target formula versions are inconsistent, but the current optimal parameter or the anchor parameter point obtains a valid target value based on the current formula retest, is not marked as a low-confidence point for evaluation, and satisfies the current parameter boundary and hardware constraints, then local candidate points are regenerated around the parameter point; otherwise, local Latin hypercube sampling is performed according to the current parameter boundary to generate local candidate points.

10. A four-point Ramsey qubit parameter initialization system, characterized in that, The method for executing the four-point Ramsey qubit parameter initialization method according to any one of claims 1-9 includes: a coarse screening sampling and parameter mapping module, a four-point Ramsey template execution module, an IQ discrimination and readout correction module, a Ramsey fidelity calculation module, a proxy optimization module, a checkpoint management module, a hardware consistency recovery module, a parameter register write-back module, and a data processing module. The coarse screening sampling and parameter mapping module, the four-point Ramsey template execution module, the IQ discrimination and readout correction module, the Ramsey fidelity calculation module, the proxy optimization module, the checkpoint management module, the hardware consistency recovery module, and the parameter register write-back module are respectively connected to the data processing module. The coarse screening and parameter mapping module is used to receive the parameter boundaries, key dimensions, fixed parameters, hardware constraints, readout assignment matrix, readout assignment matrix version, IQ discriminator version, unified four-point Ramsey experimental template version, Ramsey fidelity and optimization target formula version, and current hardware status summary of the quantum bit measurement and control parameters. It performs Latin hypercube coarse screening on the key dimensions and maps the candidate parameters to pulse configuration and readout configuration. The four-point Ramsey template execution module is used to execute a unified four-point Ramsey experimental template according to the pulse configuration and the readout configuration, under the same candidate parameters, the same hardware state represented by the current hardware state summary, the same readout assignment matrix version, and the same IQ discriminator version, and output IQ sampling data. The IQ discrimination and readout correction module is used to remove outliers, discriminate IQ data, and correct the readout assignment matrix to output four corrected excited state probabilities. The Ramsey fidelity calculation module is used to calculate short-delay visibility, long-delay visibility, and Ramsey fidelity based on the four corrected excited-state probabilities and corresponding candidate parameters. and optimization goals ; The proxy optimization module is used to construct a proxy model based on historical candidate parameters and historical optimization objectives, and to generate the next candidate parameters through the expected improvement-anchor point joint acquisition function; The checkpoint management module is used to generate optimized state checkpoints during the coarse screening and fine adjustment processes; The hardware consistency recovery module is used to perform integrity verification, unified four-point Ramsey experiment template version verification, Ramsey fidelity and optimization target formula version verification, read assignment matrix version verification, IQ discriminator version verification, hardware state summary comparison and sentinel point retesting before directly reusing the proxy model state when an interruption occurs and recovery is performed. Based on the verification results, the module controls the proxy optimization module to perform full recovery, half recovery or partial restart. The parameter register write-back module is used to write back the final optimal parameters as the initialization parameter set to the parameter register of the quantum bit measurement and control system.

Images