Intelligent ion trap voltage parameter adjusting method based on Bayesian optimization

Abstract

The invention discloses an ion trap voltage parameter intelligent adjustment method based on Bayesian optimization, and the method comprises the following steps: S1, setting a main voltage, a compensation voltage and a phase bias as to-be-optimized parameters, and forming a three-dimensional voltage parameter space; s2, generating an initial voltage combination, loading the initial voltage combination to a high-voltage power supply module, and establishing a corresponding electric field environment; s3, injecting charged particles into the ion trap in an electric field environment, and collecting a track image sequence of the particles; s4, inputting the track image sequence into an image processing module to form particle stability original data; s5, performing nonlinear state estimation on the particle stability original data in a Kalman filter, and calculating a stability evaluation index; s6, taking each group of initial voltage combination and the corresponding stability evaluation index as a sample pair, and establishing a mapping function; and S7, calling the mapping function, and determining an optimal voltage combination in combination with the joint acquisition function. According to the invention, the voltage parameters of the ion trap are intelligently adjusted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ion trap technology and precise electric field control, and particularly relates to an ion trap voltage parameter intelligent adjustment method based on Bayesian optimization. BACKGROUND

[0002] In ion trap particle manipulation experiments, accurate adjustment of voltage parameters plays a decisive role in system stability, experimental repeatability and safety. Traditional voltage adjustment usually relies on empirical rules or manual adjustment, which is not only inefficient, but also easily affected by subjective judgment of operators, and is difficult to adapt to application scenarios with complex electric field structure and significant sample differences. In recent years, with the rapid development of artificial intelligence and machine learning technology, some research has begun to introduce optimization algorithms to assist voltage configuration selection, such as genetic algorithm, particle swarm optimization, grid search, etc., to improve the intelligent level of parameter adjustment. However, these methods generally have strong assumptions on sample distribution, blind convergence process and are prone to local optimization, which makes it difficult to balance global search ability and experimental resource saving.

[0003] Especially in the face of ion trap system voltage parameter adjustment, which has the characteristics of high-dimensional input, nonlinear response and high-cost experiments, existing optimization methods often cannot fully model the complex mapping relationship between input and system stability, and lack reliable proxy models for prediction and guidance. At the same time, most methods lack modeling and constraint control of system safety boundaries in the exploration process, which can easily lead to particle drift, trap escape, and even voltage impact and other irreversible experimental accidents. In addition, existing algorithms also fail to fully consider the generality under cross-experimental conditions, and cannot effectively inherit historical optimization experience, resulting in the inability of the parameter adjustment process to migrate and continuously evolve.

[0004] Therefore, how to provide an ion trap voltage parameter intelligent adjustment method based on Bayesian optimization is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an ion trap voltage parameter intelligent adjustment method based on Bayesian optimization. The present application discloses an ion trap voltage parameter intelligent adjustment method based on Bayesian optimization, which can realize efficient and automated optimal parameter optimization in a multi-dimensional voltage parameter space. By constructing a mapping model between voltage and system stability, and combining nonlinear state estimation and Bayesian optimization strategy, the technical problems of high experimental cost, low optimization efficiency and poor parameter adjustment safety are effectively solved. This method combines image feature extraction, Kalman filter estimation and Student t process modeling, and combines maximum entropy search and constraint knowledge gradient sampling mechanism, with good adaptability and predictability. The accuracy, stability and intelligence of voltage adjustment are unified.

[0006] A method for intelligent adjustment of ion trap voltage parameters based on Bayesian optimization according to an embodiment of the present invention includes the following steps: S1. Construct a voltage control model for the quadrupole ion trap experimental system, and set the main voltage, compensation voltage and phase bias as parameters to be optimized to form a three-dimensional voltage parameter space. S2. An initial voltage combination is generated in the three-dimensional voltage parameter space using the orthogonal Latin hypercube sampling method, and each initial voltage combination is sequentially loaded onto the high-voltage power supply module to establish a stable electric field environment. S3. Based on the electric field environment, charged microparticles are injected into the ion trap, and the trajectory image sequence of the microparticles under each initial voltage combination is acquired by a sliding camera system deployed in the X and Y axes. S4. Input the acquired trajectory image sequence into the image processing module to extract feature parameters including particle drift distance, amplitude fluctuation and motion persistence to form raw particle stability data; S5. Input the raw particle stability data into the unscented transform Kalman filter for nonlinear state estimation, obtain the fused estimate of the particle state variables, and calculate the stability evaluation index. S6. Input each initial voltage combination and its corresponding stability evaluation index as a sample pair into the surrogate model constructed by the student t process. Update the surrogate model based on the trust domain rule optimization of the safe Bayesian method and establish the mapping function between voltage parameters and system performance. S7. Call the mapping function, use the joint acquisition function composed of maximum entropy search and constraint knowledge gradient to determine the optimal voltage combination in the feasible solution space, and load the voltage value into the ion trap system to complete the parameter setting and locking of the ion trap system.

[0007] Optionally, S1 specifically includes: S11. Obtain the voltage control target in the ion powder test system, and set three variables to be optimized based on the quadrupole ion trap structure. Output three voltage control parameters including main voltage, compensation voltage and phase offset angle. The main voltage is used to control the high-frequency electric field intensity, the compensation voltage is used to correct the axial symmetry of the electric field, and the phase offset angle is used to adjust the phase difference between different channel electrical signals. S12. For the three voltage control parameters, taking into account the adjustable range of the high-voltage power supply used by the system, the voltage withstand characteristics of the electrode materials, and the physical limitations of the system structure, set the adjustable range of each parameter one by one. Specifically, this includes determining the upper and lower limits of the main voltage to cover the minimum and maximum voltage output range required for system operation; setting the upper and lower limits of the compensation voltage to match the voltage difference required to adjust the electric field offset; setting the start and end values ​​of the phase offset angle to cover all possible phase configurations within a whole cycle, and outputting the value range boundaries of the three voltage control parameters. S13. Using the value range boundaries of the three voltage control parameters as input, a three-dimensional voltage parameter space is constructed through enumeration and mathematical expression to obtain a set of voltage combination points that satisfy all parameter constraints. Each point in the voltage combination point set represents a specific, legal, and usable voltage combination scheme for system input, ensuring that all optimization activities are carried out within a feasible, safe, and controlled voltage combination region.

[0008] Optionally, S2 specifically includes: S21. Read the three-dimensional voltage parameter space as input, where each sample point consists of three parameters: main voltage, compensation voltage and phase offset angle. Each parameter has a set upper and lower limit value. Set a uniform number of equal divisions for each dimension and divide the numerical range into multiple non-overlapping sub-intervals of equal width to form a discrete parameter sampling frame. S22. Based on the discrete parameter sampling frame, a position is randomly selected as a perturbation point for each sub-interval in each parameter dimension to form a Latin hypercube structure in the current dimension, ensuring that each sub-interval is visited only once. The sample sorting operation is performed independently within each dimension so that the sampling points between different dimensions are no longer directly aligned, forming an interleaved Latin sample frame. S23. Construct a sample index matrix that satisfies the modular congruence orthogonality, and group and encode the Latin sample frames in each dimension so that all sample points maintain low projection correlation in the multidimensional space and improve the orthogonality of sampling. The construction method of the sample index matrix ensures that each row represents a sample, each column corresponds to a variable dimension, and all columns are statistically orthogonal to each other. S24. Apply a nested mapping function to the sample index matrix, reconstruct the perturbation value for each dimension, and obtain the final orthogonal Latin hypercube sampling set; S25. Organize all sample points of the orthogonal Latin hypercube sampling set into an initial voltage combination according to the sampling order. Extract the main voltage value, compensation voltage value, and phase offset angle value from the initial voltage combination and use them as control inputs to load the high-voltage power controller. Apply the main voltage to the quadrupole electrode of the quadrupole ion trap through the power module. Connect the compensation voltage to the electrode plate used for vertical field distribution adjustment. At the same time, input the phase offset angle to the signal modulator for phase synchronization control between multiple high-voltage channels, establish a stable electric field environment, and apply it to the quadrupole electrode structure to trigger particle response and system performance measurement.

[0009] Optionally, S3 specifically includes: S31. Under a stable electric field environment, activate the particle delivery module in the charged particle generator, set the particle injection speed and quantity through the control system, so that the charged particles are injected into the central region of the quadrupole ion trap through the nozzle in a linear path until the particle trajectory shows a deceleration and floating trend after being captured by the electric field, and is positioned near the central axis of the trap, and output the spatial state of the charged particles under the observation state. S32. Based on the spatial state of charged particles, start the two camera units in the monitoring system, and move them to the corresponding positions of the ion trap optical window along the sliding translation system in the X and Y directions respectively. After locking the central viewpoint, start the image acquisition function synchronously, and continuously acquire particle motion images at a fixed frame rate within the set time window, and output the original data sequence of particle trajectory containing multiple frames of images. S33. Input the original data sequence of particle trajectories containing multiple frames of images into the image acquisition buffer module, and label the timestamp and current voltage combination index corresponding to each frame to realize the structured encapsulation of experimental image data into a trajectory image sequence, providing continuous, time-consistent, and identifiable data input for image processing, state estimation, and stability evaluation.

[0010] Optionally, S4 specifically includes: S41. Receive a trajectory image sequence, which records the continuous motion trajectory of charged particles under the action of an electric field, and input it to the image processing module to perform noise reduction, illumination equalization and edge enhancement on the image frames so as to clearly identify the image area of ​​charged particles in each frame and extract the centroid position of the particles in the image. S42. Combine the centroid positions of the particles in each frame in chronological order to form a complete particle position time series, which serves as the motion trajectory data corresponding to the current voltage combination. Analyze the shift of the particle position relative to the average position in the particle position time series, identify the displacement of the particle that is furthest from the center in the entire trajectory, and define it as the drift distance. This is used to measure whether the particle has a tendency to be pushed out of the stable region by the electric field. S43. Based on the particle position time series, evaluate the range of back-and-forth swing of the particles around the average position throughout the entire shooting process, calculate the overall average fluctuation of the centroid point relative to the center position as an amplitude fluctuation index to reflect whether the particles have strong periodic jitter or high-frequency oscillation behavior. S44. Based on the identification results in the particle location time series, detect whether the particles are continuously present in the identifiable area frame by frame, and count the number of continuously visible image frames. Combine the image acquisition frame rate to calculate the duration of the particles maintaining a stable state, which serves as a quantitative description of the stability of the particles under the current group voltage conditions. S45. The three parameters of drift distance, amplitude fluctuation index and duration are organized into a set of stability description indexes of charged particles and packaged into raw particle stability data.

[0011] Optionally, S5 specifically includes: S51. The original particle stability data is used as the observation input, including three parameters: the drift distance, amplitude fluctuation value and duration of existence of the particle under the current voltage combination. The state variable dimension of the unscented transform Kalman filter is set to three dimensions, which represent the steady state quantities to be estimated. The initial value of the state estimation is set to a three-dimensional vector. The state covariance matrix is ​​initialized as a diagonal positive definite matrix. The process noise covariance and observation noise covariance matrices also need to be set to reflect the system uncertainty. S52. Based on the initialized state covariance matrix, a set of sample points called sigma points are generated around the current estimated state using the unscented transformation method. These points are used to capture the probability distribution structure in the state space. Some of the sigma points are completely consistent with the current state estimate. The remaining points are constructed by adding or subtracting offsets of a certain scale in each dimension of the state vector. These points are then input into the nonlinear state transition function to simulate the state change trend under the dynamic process of the system, serving as the input basis for the prediction process. S53. Apply the nonlinear state transition function to each sigma point to obtain the set of state prediction values ​​for each sigma point at the next time step. Then, calculate the weighted reconstructed state prediction covariance based on the weighted average of the set of state prediction values ​​for all sigma points. S54. The set of state prediction values ​​is compared with the observation data at the current time, namely the three indicators of drift, amplitude and duration from the image processing stage, to construct the observation prediction error. The cross covariance between the predicted state and the observation is calculated. Then, based on the ratio between the cross covariance and the observation prediction error, the Kalman gain at the current time is obtained. The Kalman gain is used to correct the state prediction, thereby updating the fusion estimate of the particle state quantity. The fusion estimate integrates the dual information of the prediction model and the actual observation. S55. Extract particle drift estimate, amplitude estimate and duration estimate from the fusion estimate of particle state variables, and calculate the stability evaluation index by weighted combination. The stability evaluation index retains the comprehensive characteristics of particle state in structure, serves as the performance evaluation result of particle response under the current voltage combination, and is used to guide the update of the objective function and the adjustment of sampling direction in Bayesian optimization.

[0012] 7. The intelligent adjustment method for ion trap voltage parameters based on Bayesian optimization according to claim 1, wherein step S6 specifically includes: S61. Each initial voltage combination generated in the experimental stage is represented as a three-dimensional input item in vector form, which includes the main voltage value, the compensation voltage value and the modulation phase angle. The three-dimensional input item is paired with the corresponding stability evaluation index to form an input-output sample pair. S62. Construct a student t-process proxy model. The output of the student t-process proxy model is set to follow a multivariate student t-distribution, where the prior distribution is defined by a mean function and a covariance kernel function. The Matérn5 / 2 kernel function is used to adapt to the non-stationarity and outliers present in the voltage regulation scenario, and the covariance function is output. The expression of the Matérn5 / 2 kernel function is: ; in, , These represent any two voltage combinations. The covariance function measures the correlation between two voltage combinations in prediction. This represents the Euclidean distance between any two voltage input samples. For length scale hyperparameters, The output variance coefficients are used to control the amplitude range of the model's output response. It is an exponential function with the natural constant as its base; S63. When a new voltage combination is input, the student t-process proxy model will use the training sample data to predict the posterior output and obtain a prediction result of the student t-distribution with triple information, including the predicted mean of the input point under the current model, the prediction variance of the prediction uncertainty of the current point, and a degree of freedom parameter representing the thickness of the distribution tail. The prediction result is calculated based on the weighted combination of the covariances between all training points and the current input. The prediction variance is determined by the covariance structure between the current point and the training points and the inverse matrix of the covariance matrix of the training data, and a degree of freedom adjustment term is added to correct the overall uncertainty. S64. To ensure that voltage sampling is always within the safe operating range of the system during the optimization process, a trust domain strategy for safe Bayesian optimization is defined. The predicted mean and predicted variance of the new input point are calculated. The value of the acquisition function is constructed by adding a confidence factor adjusted according to the degree of freedom parameter to the predicted mean and multiplying it with the predicted variance. The model is restricted to selecting input points that meet the condition that the predicted mean minus the confidence interval is still higher than a certain stability safety threshold to enter the experiment. Based on the value of the acquisition function, only points whose lower confidence bound exceeds the preset safety value are allowed to participate in the next round of voltage combination evaluation, forming a trust domain constraint, and obtaining sampling candidate points that meet the trust domain constraint. S65. Select the input point with the largest covariance function value from the sampled candidate points that satisfy the trust domain constraint, and use it as the voltage recommendation combination for the current round of Bayesian optimization process. Input the voltage recommendation combination into the actual experimental process, collect voltage parameters and stability indexes, add them to the training set as new sample pairs, and update and iterate the student t process model to gradually establish the mapping function between voltage parameters and system performance.

[0013] Optionally, S7 specifically includes: S71. Construct a maximum entropy search acquisition function to measure the contribution of each candidate input point to the estimation of the optimal value of the objective function in terms of information entropy, and output the maximum entropy acquisition value. S72. Based on the maximum entropy collected value, in order to take into account the long-term benefit optimization goal, a constrained knowledge gradient strategy is introduced for supplementary evaluation. The maximum predicted mean of all feasible input points is estimated when the sample set is not updated, and the model is updated once based on the currently selected point to estimate the new maximum predicted mean. The difference between the maximum predicted mean and the new maximum predicted mean is the knowledge gradient value, which can be used to guide the strategy to tilt towards long-term benefits. S73. A joint sampling function is constructed by weighted fusion, which combines the maximum entropy sampling value and the knowledge gradient value. A weight factor is set to adjust the contribution of the two. The maximum entropy sampling value and the knowledge gradient value are linearly superimposed in proportion to obtain the joint sampling score of each input point, which reflects the comprehensive superiority of the current input point in both information gain and long-term adjustment benefits. S74. In the feasible solution space constrained by the trust domain, traverse all candidate voltage combination input points in turn and calculate the corresponding joint sampling score. Select the voltage combination point with the largest joint acquisition function value as the optimal voltage combination recommended by Bayesian optimization in the current round, and load the voltage value into the ion trap system to update the system control parameters and trigger the next experiment.

[0014] The beneficial effects of this invention are: 1. Compared with traditional parameter tuning methods that rely on experience or simple optimization algorithms, this method introduces a student t-process surrogate model, which can more accurately capture the nonlinear mapping relationship between voltage parameters and particle stability, significantly improving the model's predictive and generalization capabilities.

[0015] 2. By fusing and estimating the particle trajectory features extracted from the image through an unscented transform Kalman filter, the state fluctuations caused by system noise and measurement errors are effectively suppressed, enhancing the reliability and physical interpretability of the stability index.

[0016] 3. The joint acquisition function, which combines maximum entropy search and constrained knowledge gradient construction, not only improves the global optimality of parameter search, but also effectively avoids high-risk experimental states such as voltage jumps and particle decoupling by introducing a safe Bayesian optimization trust domain mechanism, thus ensuring the safety and stability of the optimization process.

[0017] 4. Finally, this method possesses excellent cross-batch transfer capability and a feedback closed-loop structure, enabling the accumulation of parameter tuning experience through multiple rounds of experiments, achieving continuous iteration and self-evolution of the intelligent adjustment process. Overall, this invention significantly improves the efficiency and control precision of ion trap experiments while ensuring adjustment safety, and has broad application prospects and practical promotion value. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Fig. 1 This is a flowchart of an intelligent adjustment method for ion trap voltage parameters based on Bayesian optimization proposed in this invention; Fig. 2 This is a schematic diagram of the algorithm structure of an intelligent adjustment method for ion trap voltage parameters based on Bayesian optimization proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figs. 1-2 A method for intelligent adjustment of ion trap voltage parameters based on Bayesian optimization includes the following steps: S1. Construct a voltage control model for the quadrupole ion trap experimental system, and set the main voltage, compensation voltage and phase bias as parameters to be optimized to form a three-dimensional voltage parameter space. S2. An initial voltage combination is generated in the three-dimensional voltage parameter space using the orthogonal Latin hypercube sampling method, and each initial voltage combination is sequentially loaded onto the high-voltage power supply module to establish a stable electric field environment. S3. Based on the electric field environment, charged microparticles are injected into the ion trap, and the trajectory image sequence of the microparticles under each initial voltage combination is acquired by a sliding camera system deployed in the X and Y axes. S4. Input the acquired trajectory image sequence into the image processing module to extract feature parameters including particle drift distance, amplitude fluctuation and motion persistence to form raw particle stability data; S5. Input the raw particle stability data into the unscented transform Kalman filter for nonlinear state estimation, obtain the fused estimate of the particle state variables, and calculate the stability evaluation index. S6. Input each initial voltage combination and its corresponding stability evaluation index as a sample pair into the surrogate model constructed by the student t process. Update the surrogate model based on the trust domain rule optimization of the safe Bayesian method and establish the mapping function between voltage parameters and system performance. S7. Call the mapping function, use the joint acquisition function composed of maximum entropy search and constraint knowledge gradient to determine the optimal voltage combination in the feasible solution space, and load the voltage value into the ion trap system to complete the parameter setting and locking of the ion trap system.

[0022] This invention generates experimental combinations in a three-dimensional voltage parameter space through orthogonal Latin hypercube sampling, combined with image feature extraction and unscented transform Kalman filtering to accurately obtain particle stability indices. Based on this, relying on a Student's t-process surrogate model and a secure Bayesian optimization trust domain strategy, it achieves modeling of the mapping between voltage and system performance. Furthermore, through a joint sampling mechanism of maximum entropy search and constrained knowledge gradients, it intelligently makes decisions to achieve globally optimal parameter selection and stable system operation.

[0023] In this embodiment, S1 specifically includes: S11. Obtain the voltage control target in the ion powder test system, and set three variables to be optimized based on the quadrupole ion trap structure. Output three voltage control parameters including main voltage, compensation voltage and phase offset angle. The main voltage is used to control the high-frequency electric field intensity, the compensation voltage is used to correct the axial symmetry of the electric field, and the phase offset angle is used to adjust the phase difference between different channel electrical signals. S12. For the three voltage control parameters, taking into account the adjustable range of the high-voltage power supply used by the system, the voltage withstand characteristics of the electrode materials, and the physical limitations of the system structure, set the adjustable range of each parameter one by one. Specifically, this includes determining the upper and lower limits of the main voltage to cover the minimum and maximum voltage output range required for system operation; setting the upper and lower limits of the compensation voltage to match the voltage difference required to adjust the electric field offset; setting the start and end values ​​of the phase offset angle to cover all possible phase configurations within a whole cycle, and outputting the value range boundaries of the three voltage control parameters. S13. Using the value range boundaries of the three voltage control parameters as input, a three-dimensional voltage parameter space is constructed through enumeration and mathematical expression to obtain a set of voltage combination points that satisfy all parameter constraints. Each point in the voltage combination point set represents a specific, legal, and usable voltage combination scheme for system input, ensuring that all optimization activities are carried out within a feasible, safe, and controlled voltage combination region.

[0024] This step involves extracting three key control parameters—main voltage, compensation voltage, and phase bias angle—to establish a high-dimensional adjustable parameter model covering the system's operational requirements. By considering the high-voltage power supply performance, electrode material voltage withstand limitations, and structural physical constraints, clear upper and lower limits are set for each control parameter to ensure the safety and effectiveness of parameter adjustment. Based on this, a three-dimensional voltage parameter space is constructed, forming a set of voltage combination points that satisfy the constraints. This serves as the input basis for subsequent optimization and adjustment strategies, providing reliable physical and engineering guarantees for the overall system's controllability, robustness, and experimental stability.

[0025] In this embodiment, S2 specifically includes: S21. Read the three-dimensional voltage parameter space as input, where each sample point consists of three parameters: main voltage, compensation voltage and phase offset angle. Each parameter has a set upper and lower limit value. Set a uniform number of equal divisions for each dimension and divide the numerical range into multiple non-overlapping sub-intervals of equal width to form a discrete parameter sampling frame. S22. Based on the discrete parameter sampling frame, a position is randomly selected as a perturbation point for each sub-interval in each parameter dimension to form a Latin hypercube structure in the current dimension, ensuring that each sub-interval is visited only once. The sample sorting operation is performed independently within each dimension so that the sampling points between different dimensions are no longer directly aligned, forming an interleaved Latin sample frame. S23. Construct a sample index matrix that satisfies the modular congruence orthogonality, and group and encode the Latin sample frames in each dimension so that all sample points maintain low projection correlation in the multidimensional space and improve the orthogonality of sampling. The construction method of the sample index matrix ensures that each row represents a sample, each column corresponds to a variable dimension, and all columns are statistically orthogonal to each other. S24. Apply a nested mapping function to the sample index matrix, reconstruct the perturbation value for each dimension, and obtain the final orthogonal Latin hypercube sampling set. The nested mapping function generates the numerical expression of the sample in each sampling dimension as follows: ; in, Indicates the first The sample at the th The actual values ​​of each parameter dimension and These are the lower and upper limits of the current dimension, respectively. This refers to the arrangement number of the current sampling point within the layer after the disturbance. This is a random perturbation term between zero and one, used to introduce fine-tuning at sampling points; S25. Organize all sample points of the orthogonal Latin hypercube sampling set into an initial voltage combination according to the sampling order. Extract the main voltage value, compensation voltage value, and phase offset angle value from the initial voltage combination and use them as control inputs to load the high-voltage power controller. Apply the main voltage to the quadrupole electrode of the quadrupole ion trap through the power module. Connect the compensation voltage to the electrode plate used for vertical field distribution adjustment. At the same time, input the phase offset angle to the signal modulator for phase synchronization control between multiple high-voltage channels, establish a stable electric field environment, and apply it to the quadrupole electrode structure to trigger particle response and system performance measurement.

[0026] This step achieves efficient sampling of the main voltage, compensation voltage, and phase bias angle within a three-dimensional voltage parameter space by constructing an orthogonal Latin hypercube sampling mechanism with low correlation and high uniformity. Statistical independence among multidimensional variables is ensured by constructing a modulo-congruent orthogonal index matrix. Based on this, a nested perturbation mapping function is applied to introduce fine-tuning perturbations into the sampling points, generating a spatially uniform and staggered orthogonal Latin hypercube sample set. All sampling points are then converted into control inputs and loaded into the high-voltage power supply control module, effectively constructing a stable and reliable electric field environment and providing comprehensive and structured initial experimental conditions for subsequent particle response acquisition and performance evaluation.

[0027] In this embodiment, S3 specifically includes: S31. Under a stable electric field environment, activate the particle delivery module in the charged particle generator, set the particle injection speed and quantity through the control system, so that the charged particles are injected into the central region of the quadrupole ion trap through the nozzle in a linear path until the particle trajectory shows a deceleration and floating trend after being captured by the electric field, and is positioned near the central axis of the trap, and output the spatial state of the charged particles under the observation state. S32. Based on the spatial state of charged particles, start the two camera units in the monitoring system, and move them to the corresponding positions of the ion trap optical window along the sliding translation system in the X and Y directions respectively. After locking the central viewpoint, start the image acquisition function synchronously, and continuously acquire particle motion images at a fixed frame rate within the set time window, and output the original data sequence of particle trajectory containing multiple frames of images. S33. Input the original data sequence of particle trajectories containing multiple frames of images into the image acquisition buffer module, and label the timestamp and current voltage combination index corresponding to each frame to realize the structured encapsulation of experimental image data into a trajectory image sequence, providing continuous, time-consistent, and identifiable data input for image processing, state estimation, and stability evaluation.

[0028] This step utilizes a sliding camera system deployed along the X and Y axes to continuously observe the spatial behavior of particles from multiple angles, acquiring multi-frame trajectory image data containing temporal information and spatial dynamic characteristics. All image sequences are labeled with time and voltage indices frame-by-frame in the acquisition buffer module, completing the structured encapsulation of the data. This provides an accurate, continuous, and time-consistent input foundation for subsequent image feature extraction and stability analysis, ensuring the system's dynamic perception capability of voltage regulation effects.

[0029] In this embodiment, S4 specifically includes: S41. Receive a trajectory image sequence, which records the continuous motion trajectory of charged particles under the action of an electric field, and input it to the image processing module to perform noise reduction, illumination equalization and edge enhancement on the image frames so as to clearly identify the image area of ​​charged particles in each frame and extract the centroid position of the particles in the image. S42. Combine the centroid positions of the particles in each frame in chronological order to form a complete particle position time series, which serves as the motion trajectory data corresponding to the current voltage combination. Analyze the shift of the particle position relative to the average position in the particle position time series, identify the displacement of the particle that is furthest from the center in the entire trajectory, and define it as the drift distance. This is used to measure whether the particle has a tendency to be pushed out of the stable region by the electric field. S43. Based on the particle position time series, evaluate the range of back-and-forth swing of the particles around the average position throughout the entire shooting process, calculate the overall average fluctuation of the centroid point relative to the center position as an amplitude fluctuation index to reflect whether the particles have strong periodic jitter or high-frequency oscillation behavior. S44. Based on the identification results in the particle location time series, detect whether the particles are continuously present in the identifiable area frame by frame, and count the number of continuously visible image frames. Combine the image acquisition frame rate to calculate the duration of the particles maintaining a stable state, which serves as a quantitative description of the stability of the particles under the current group voltage conditions. S45. The three parameters of drift distance, amplitude fluctuation index and duration are organized into a set of stability description indexes of charged particles and packaged into raw particle stability data.

[0030] This step uses a sequence of trajectory images as a foundation to perform structured analysis of particle motion behavior through an image processing module, extracting key physical feature parameters. After denoising and enhancement, the centroid positions of the particles are clearly identified, generating a complete position-time series. This series is used to calculate the maximum particle offset distance, average amplitude fluctuation, and duration of identifiable movement, corresponding to core indicators such as drift trend, oscillation intensity, and stable existence, respectively. These features collectively constitute a quantitative expression of the particle's dynamic response under voltage combinations, providing accurate and comparable raw stability data to support the system's evaluation of electric field stability and the controlled state of the particles.

[0031] In this embodiment, S5 specifically includes: S51. The original particle stability data is used as the observation input, including three parameters: the drift distance, amplitude fluctuation value and duration of existence of the particle under the current voltage combination. The state variable dimension of the unscented transform Kalman filter is set to three dimensions, which represent the steady state quantities to be estimated. The initial value of the state estimation is set to a three-dimensional vector. The state covariance matrix is ​​initialized as a diagonal positive definite matrix. The process noise covariance and observation noise covariance matrices also need to be set to reflect the system uncertainty. S52. Based on the initialized state covariance matrix, a set of sample points called sigma points are generated around the current estimated state using the unscented transformation method. These points are used to capture the probability distribution structure in the state space. Some of the sigma points are completely consistent with the current state estimate. The remaining points are constructed by adding or subtracting offsets of a certain scale in each dimension of the state vector. These points are then input into the nonlinear state transition function to simulate the state change trend under the dynamic process of the system, serving as the input basis for the prediction process. S53. Apply the nonlinear state transition function to each sigma point to obtain the set of state prediction values ​​for each sigma point at the next time step. Then, calculate the weighted reconstructed state prediction covariance based on the weighted average of the state prediction value sets of all sigma points. The calculation expression is as follows: ; in, Indicates the current time State prediction covariance matrix Indicates time The One predicted sigma point, Indicates time The mean of the state prediction, Indicates the first Covariance weighting coefficients corresponding to each sigma point The process noise covariance matrix; S54. The set of state prediction values ​​is compared with the observation data at the current time, namely the three indicators of drift, amplitude and duration from the image processing stage, to construct the observation prediction error. The cross covariance between the predicted state and the observation is calculated. Then, based on the ratio between the cross covariance and the observation prediction error, the Kalman gain at the current time is obtained. The Kalman gain is used to correct the state prediction, thereby updating the fusion estimate of the particle state quantity. The fusion estimate integrates the dual information of the prediction model and the actual observation. S55. Extract particle drift estimate, amplitude estimate and duration estimate from the fusion estimate of particle state variables, and calculate the stability evaluation index by weighted combination. The stability evaluation index retains the comprehensive characteristics of particle state in structure, serves as the performance evaluation result of particle response under the current voltage combination, and is used to guide the update of the objective function and the adjustment of sampling direction in Bayesian optimization.

[0032] This step initializes state variables and covariance matrices to construct a state space and generate a set of sigma points representing the system's distribution characteristics, used to simulate the changing trends of the state during dynamic evolution. The observation data obtained in the image processing stage is fused, and the prediction results are adjusted using Kalman gain to achieve an effective combination of prediction and observation, outputting a fused estimate that more closely approximates the actual state. Finally, core stability parameters are extracted from the fused estimate, and stability evaluation indices are calculated, providing a reliable performance benchmark for surrogate model training and sampling strategies, improving the physical accuracy and regulatory feedback efficiency of the parameter optimization process.

[0033] In this embodiment, S6 specifically includes: S61. Each initial voltage combination generated in the experimental stage is represented as a three-dimensional input item in vector form, which includes the main voltage value, the compensation voltage value and the modulation phase angle. The three-dimensional input item is paired with the corresponding stability evaluation index to form an input-output sample pair. S62. Construct a student t-process proxy model. The output of the student t-process proxy model is set to follow a multivariate student t-distribution, where the prior distribution is defined by a mean function and a covariance kernel function. The Matérn5 / 2 kernel function is used to adapt to the non-stationarity and outliers present in the voltage regulation scenario, and the covariance function is output. The expression of the Matérn5 / 2 kernel function is: ; in, , These represent any two voltage combinations. The covariance function measures the correlation between two voltage combinations in prediction. This represents the Euclidean distance between any two voltage input samples. For length scale hyperparameters, The output variance coefficients are used to control the amplitude range of the model's output response. It is an exponential function with the natural constant as its base; S63. When a new voltage combination is input, the student t-process proxy model will use the training sample data to predict the posterior output and obtain a prediction result of the student t-distribution with triple information, including the predicted mean of the input point under the current model, the prediction variance of the prediction uncertainty of the current point, and a degree of freedom parameter representing the thickness of the distribution tail. The prediction result is calculated based on the weighted combination of the covariances between all training points and the current input. The prediction variance is determined by the covariance structure between the current point and the training points and the inverse matrix of the covariance matrix of the training data, and a degree of freedom adjustment term is added to correct the overall uncertainty. S64. To ensure that voltage sampling is always within the safe operating range of the system during the optimization process, a trust domain strategy for safe Bayesian optimization is defined. The predicted mean and predicted variance of the new input point are calculated. The value of the acquisition function is constructed by adding a confidence factor adjusted according to the degree of freedom parameter to the predicted mean and multiplying it with the predicted variance. The model is restricted to selecting input points that meet the condition that the predicted mean minus the confidence interval is still higher than a certain stability safety threshold to enter the experiment. Based on the value of the acquisition function, only points whose lower confidence bound exceeds the preset safety value are allowed to participate in the next round of voltage combination evaluation, forming a trust domain constraint, and obtaining sampling candidate points that meet the trust domain constraint. S65. Select the input point with the largest covariance function value from the sampled candidate points that satisfy the trust domain constraint, and use it as the voltage recommendation combination for the current round of Bayesian optimization process. Input the voltage recommendation combination into the actual experimental process, collect voltage parameters and stability indexes, add them to the training set as new sample pairs, and update and iterate the student t process model to gradually establish the mapping function between voltage parameters and system performance.

[0034] This step models the non-stationary input using the Matérn5 / 2 covariance kernel function, enhancing its adaptability to anomalous responses to discrete voltage combinations. Based on this, the model's predicted output includes three indices: mean, variance, and degrees of freedom, comprehensively reflecting the response estimate and its uncertainty for the current input. Furthermore, a safe Bayesian optimization trust region mechanism is introduced, selecting sampling points within the safe boundary based on the confidence interval to ensure the reliability of the experiment and system protection. Finally, the experimental feedback is performed using the candidate point with the largest covariance, achieving a closed-loop iteration between the model and the system response, providing a highly robust mapping function support for voltage parameter optimization.

[0035] In this embodiment, S7 specifically includes: S71. Construct a maximum entropy search acquisition function to measure the contribution of each candidate input point to the estimation of the optimal value of the objective function in terms of information entropy, and output the maximum entropy acquisition value. The expression of the maximum entropy search acquisition function is as follows: ; in, As candidate input points, This is the maximum entropy value collected. This represents the current training sample set. Indicates to The predicted value represents the predicted distribution of the maximum value of the objective function. Indicates the predicted extreme value. To calculate the expectation of the global extreme value prediction distribution, To calculate the expected response for the input point, Let entropy be the response to the current point under the condition of predicting the extreme value. Given the current sample set, the larger the output value of the information entropy function, the more likely the current point is to improve the perception of the global optimum; S72. Based on the maximum entropy collected value, in order to take into account the long-term benefit optimization goal, a constrained knowledge gradient strategy is introduced for supplementary evaluation. The maximum predicted mean of all feasible input points is estimated when the sample set is not updated, and the model is updated once based on the currently selected point to estimate the new maximum predicted mean. The difference between the maximum predicted mean and the new maximum predicted mean is the knowledge gradient value, which can be used to guide the strategy to tilt towards long-term benefits. S73. A joint sampling function is constructed by weighted fusion, which combines the maximum entropy sampling value and the knowledge gradient value. A weight factor is set to adjust the contribution of the two. The maximum entropy sampling value and the knowledge gradient value are linearly superimposed in proportion to obtain the joint sampling score of each input point, which reflects the comprehensive superiority of the current input point in both information gain and long-term adjustment benefits. S74. In the feasible solution space constrained by the trust domain, traverse all candidate voltage combination input points in turn and calculate the corresponding joint sampling score. Select the voltage combination point with the largest joint acquisition function value as the optimal voltage combination recommended by Bayesian optimization in the current round, and load the voltage value into the ion trap system to update the system control parameters and trigger the next experiment.

[0036] This step combines maximum entropy search and constrained knowledge gradient strategies to construct a joint acquisition mechanism that integrates information gain and long-term optimization benefits. This mechanism is used to intelligently select the optimal parameter combination within the feasible voltage space. The maximum entropy acquisition function measures the potential of candidate points to improve the prediction of the global optimum, capturing information gaps in the current model. The knowledge gradient strategy evaluates the long-term improvement effect of candidate points on the predictive ability of future surrogate models. The two are weighted and fused to form a joint acquisition function that balances exploratory and robustness, ensuring that the adjustment process is both innovative and maintains control stability. Finally, based on the joint score, the optimal voltage combination is selected within the trust domain for system loading, achieving efficient, safe, and closed-loop experimental optimization iterations.

[0037] Example 1: To verify the feasibility of this invention in practice, it was applied to a particle stability control experiment in an ion trap system at a precision particle manipulation laboratory. This experiment involved the capture and retention of charged particles with high sensitivity and safety requirements, involving multiple variables. The experimental objective was to rapidly find the optimal voltage combination within a three-dimensional voltage parameter space to achieve long-term, stable, and low-drift capture control of charged particles in a quadrupole ion trap. The quadrupole ion trap structure used in the experiment included one set of main control electrodes, two sets of compensation and adjustment electrodes, and a multi-channel high-voltage signal synchronization module. The system control voltage included the main voltage, compensation voltage, and modulation phase bias angle. To optimize the voltage parameters of this system, the Bayesian optimization-based intelligent adjustment method described in this invention was employed.

[0038] First, a voltage control model was constructed, setting the main voltage range to 280–400V, the compensation voltage range to -35–35V, and the phase offset angle range to 0–360°. Using the orthogonal Latin hypercube sampling method, 60 initial voltage combinations were generated in the three-dimensional parameter space and loaded into the system for capture experiments. Under each voltage condition, the system acquired particle trajectory images along the X and Y axes using a sliding camera, extracting three stability characteristics: maximum drift distance, mean amplitude fluctuation, and duration of existence.

[0039] The acquired image sequence data, after being processed by image processing and an unscented transform Kalman filter, is transformed into a stability evaluation index for training the student process surrogate model. During sampling and prediction, the model incorporates the Matérn 5 / 2 covariance kernel function to characterize the nonlinear correlation between input variables, eliminating sampling points in potentially unstable regions. In seven rounds of Bayesian optimization iterations, a joint acquisition function is introduced in each round, integrating a strategy of maximum entropy search and constrained knowledge gradients for sampling decisions. Finally, in the sixth round, the voltage combination with the optimal stability index is converged: main voltage 354V, compensation voltage –8.5V, and phase offset angle 225°.

[0040] Table 1. Experimental comparison between the present invention and traditional methods in particle stability regulation.

[0041] As can be seen from the experimental comparison table 1, the system using the method of this invention shows significant improvements over the traditional grid search parameter tuning method in terms of average drift control, stability maintenance time, and response consistency. In the average 3 hours of system operation, the maximum particle drift based on the parameter tuning scheme of this invention is reduced from 27.6 μm under the traditional method to 9.3 μm, and the particle stability maintenance time is significantly improved from 21.3s to 46.7s, and no de-trajectory or trajectory loss phenomena are observed during the process.

[0042] In precisely controlled environments, this invention achieves efficient, safe, and stable parameter adjustment performance, significantly outperforming traditional methods and validating its application value in ion trap system optimization. This method can significantly shorten the experimental parameter tuning cycle, reduce human intervention costs, and effectively avoid experimental risks caused by voltage surges and unstable parameter tuning, demonstrating good engineering deployability and research promotion value.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent adjustment of ion-trap voltage parameters based on Bayesian optimization, characterized in that, The method comprises the following steps: S1, constructing a voltage control model of a quadrupole ion trap experimental system, and setting a main voltage, a compensation voltage and a phase bias as to-be-optimized parameters to form a three-dimensional voltage parameter space; S2, generating initial voltage combinations in the three-dimensional voltage parameter space by using an orthogonal Latin hypercube sampling method, and sequentially loading each initial voltage combination to a high-voltage power supply module to establish a stable loaded electric field environment; S3, injecting charged particles into the ion trap based on the electric field environment, and collecting trajectory image sequences of the particles under each initial voltage combination by a sliding camera system arranged in the X-axis and Y-axis directions; S4, inputting the collected trajectory image sequences into an image processing module to extract feature parameters including particle drift distance, amplitude fluctuation and motion duration to form particle stability original data; S5, inputting the particle stability original data into an unscented Kalman filter for nonlinear state estimation to obtain fused estimates of particle state quantities and calculate a stability evaluation index; S6, inputting each initial voltage combination and the corresponding stability evaluation index as a sample pair into a surrogate model constructed based on a student t process, updating the surrogate model based on a safe Bayesian optimization trust region rule, and establishing a mapping function between the voltage parameters and the system performance; S7, calling the mapping function, determining an optimal voltage combination in a feasible solution space by combining a joint collection function formed by a maximum entropy search and a constraint knowledge gradient, and loading the voltage values to the ion trap system.

2. The method of claim 1, wherein, The S1 specifically comprises: S11, obtaining voltage control targets in an ion powder test system, setting three variables to be optimized based on a quadrupole ion trap structure, and outputting three voltage control parameters including a main voltage, a compensation voltage and a phase bias angle; S12, setting the adjustable range of each parameter one by one for the three voltage control parameters in combination with the adjustable range of the high-voltage power supply used by the system, the voltage resistance characteristics of the electrode material and the physical limitations of the system structure, and outputting the value range boundaries of the three voltage control parameters; S13, taking the value range boundaries of the three voltage control parameters as input, and constructing a three-dimensional voltage parameter space through enumeration and mathematical expression.

3. The method of claim 1, wherein, The S2 specifically comprises: S21, reading the three-dimensional voltage parameter space as input, setting a uniform number of divisions for each dimension, dividing the numerical range into multiple non-overlapping and equal-width subintervals, and forming a discrete parameter sampling frame; S22, based on the discrete parameter sampling frame, randomly extracting a position as a perturbation point for each subinterval in each parameter dimension to form a Latin hypercube structure in the current dimension, and independently performing sample sorting operations within each dimension to form an interleaved Latin sample framework; S23, constructing a sample index matrix satisfying modular congruence orthogonality, and grouping and encoding the Latin sample framework in each dimension; S24, applying a nested mapping function to the sample index matrix, reconstructing the perturbation value for each dimension, and obtaining a final orthogonal Latin hypercube sampling set. S25, all sample points of the orthogonal Latin hypercube sampling set are organized into an initial voltage combination according to a sampling sequence, a main voltage value, a compensation voltage value and a phase offset angle value in the initial voltage combination are extracted, and a stable loaded electric field environment is established.

4. The method of claim 1, wherein, The S3 specifically includes: S31, under the stable loaded electric field environment, a particle spraying module in the charged particle generator is activated, the particle spraying speed and quantity are set by the control system, and the charged particle space state in the to-be-observed state is output; S32, based on the charged particle space state, two camera units in the monitoring system are started, and the particle motion images are continuously collected at a fixed frame rate within a set time window, and a particle trajectory original data sequence containing multiple images is output; S33, the particle trajectory original data sequence containing multiple images is input into an image collection buffer module, and each frame corresponding time stamp and current voltage combination index are labeled, and a trajectory image sequence is packaged.

5. The method of claim 1, wherein, The S4 specifically includes: S41, the trajectory image sequence is received and input into an image processing module, noise removal processing, light balance and edge enhancement are performed on the image frames, and the particle centroid position in the image is extracted; S42, the particle centroid position of each frame obtained is combined in time sequence to form a complete particle position time sequence, and the offset change of the particle position relative to the average position in the particle position time sequence is analyzed, the displacement of the particle deviating from the center farthest in the entire trajectory is identified, and is defined as the drift distance; S43, based on the particle position time sequence, the swing range of the particle around the average position in the entire shooting process is evaluated, the overall fluctuation mean of the centroid point to the center position is calculated, and is taken as the amplitude fluctuation index; S44, relying on the identification result in the particle position time sequence, the number of continuously visible image frames is counted, and the duration length of the particle maintaining the stable state is calculated in combination with the image collection frame rate; S45, the drift distance, the amplitude fluctuation index and the duration length are uniformly organized into a set of stability description indexes of the charged particle, and are packaged as particle stability original data.

6. The method of claim 1, wherein, The S5 specifically includes: S51, the particle stability original data is taken as an observation input, the state variable dimension of the unscented transformation Kalman filter is set to three dimensions, and the state covariance matrix is initialized as a diagonal positive definite matrix; S52, according to the initialized state covariance matrix, a set of sample points called sigma points are generated around the current estimated state by using the unscented transformation method, and are input into a nonlinear state transition function to simulate the trend of state change under the system dynamic process; S53, the nonlinear state transition function is applied to the sigma points respectively, a set of state prediction values of each sigma point at the next time is obtained, and then the weighted reconstruction state prediction covariance is calculated according to the weighted average value of all sigma point state prediction value sets; S54, the state prediction value set is compared with the observation data at the current time, the observation prediction error is constructed, the cross covariance between the prediction state and the observation is calculated, the Kalman gain at the current time is obtained, and the fusion estimation value of the particle state quantity is obtained by updating. S55, extract the particle drift estimation, amplitude estimation and persistence time estimation from the fusion estimation of the particle state quantity respectively, and calculate the stability evaluation index by weighted combination.

7. The method of claim 1, wherein the method is based on Bayesian optimization of ion trap voltage parameters. The S6 specifically comprises: S61, each group of initial voltage combinations generated in the experimental stage is paired with the corresponding stability evaluation index one by one to form an input-output sample pair; S62, a student t process proxy model is constructed, the output of the student t process proxy model is set to obey a multivariate student t distribution, a Matérn 5 / 2 kernel function is adopted to adapt to the non-stationarity and outliers existing in the voltage regulation scene, and a covariance function is output; S63, when a new voltage combination is input, the student t process proxy model will predict the posterior output using the training sample data to obtain a prediction result of the student t distribution with triple information, including the prediction mean of the input point under the current model, the prediction variance of the prediction uncertainty of the current point, and a degree of freedom parameter representing the thickness of the distribution tail; S64, define the trust region strategy of safe Bayesian optimization, calculate the prediction mean and prediction variance of the new input point, and form the value of the acquisition function by adding a confidence factor adjusted according to the degree of freedom parameter to the prediction mean and multiplying the prediction variance, based on the value of the acquisition function, only the points with the lower confidence limit exceeding the preset safety value are allowed to participate in the next round of voltage combination evaluation, forming a trust region constraint, and obtaining a sampling candidate point that satisfies the trust region constraint; S65, select the input point with the maximum covariance function value from the sampling candidate points that satisfy the trust region constraint, collect the voltage parameters and stability index, and add them to the training set as new sample pairs, and update and iterate the student t process model to gradually establish the mapping function between the voltage parameters and the system performance.

8. The method of claim 1, wherein, The S7 specifically comprises: S71, construct a maximum entropy search acquisition function to measure the contribution of each candidate input point to the estimation of the optimal value of the objective function in the sense of information entropy, and output the maximum entropy acquisition value; S72, based on the maximum entropy acquisition value, introduce a constraint knowledge gradient strategy for supplementary evaluation, estimate the maximum prediction mean of all feasible input points without updating the sample set, and update the model based on the currently selected point to estimate the new maximum prediction mean, and the difference between the maximum prediction mean and the new maximum prediction mean is the knowledge gradient value; S73, construct a joint acquisition function by weighted fusion, integrate the maximum entropy acquisition value and the knowledge gradient value, set a weight factor, linearly superimpose the maximum entropy acquisition value and the knowledge gradient value in proportion, and obtain the joint sampling score of each input point; S74, traverse all candidate voltage combination input points in the feasible solution space constrained by the trust region in turn, and calculate the corresponding joint sampling score, and select the voltage combination point with the maximum joint acquisition function value as the optimal voltage combination recommended by the current round of Bayesian optimization, and load the voltage value to the ion trap system.

Images