A neural network-based aod optical tweezer array automatic correction method and system

Abstract

The application belongs to the field of quantum physics and quantum technology, and particularly relates to an AOD optical tweezer array automatic correction method and system based on a neural network, comprising the following steps: generating an initial AOD optical tweezer array through an acousto-optic deflector; collecting an image of the initial AOD optical tweezer array and pre-processing the image; inputting the pre-processed image into a first neural network and a second neural network; calculating a global frequency compensation amount according to the average position deviation and a pre-calibrated frequency-position relationship factor; generating an intensity compensation table covering the whole frequency domain through an interpolation method based on the intensity gain compensation factor of each optical tweezer; updating the driving frequency of the AOD according to the global frequency compensation amount, and updating the driving amplitude of each channel of the AOD according to the intensity compensation table, and outputting the corrected AOD optical tweezer array. Through the integration of the neural network model and the closed-loop control process, the position and intensity correction of the optical tweezer array is fully automated.

Description

Technical Field

[0001] This invention belongs to the field of quantum physics and quantum technology, and in particular relates to an automatic calibration method and system for AOD optical tweezers array based on neural networks. Background Technology

[0002] Optical tweezers utilize focused lasers to generate gradient forces at the microscopic scale, enabling the capture and manipulation of atoms, molecules, and particles. In fields such as cold atom physics, quantum information processing, and ultracold molecular assembly, optical tweezers arrays have become an important technology for constructing large-scale qubit systems.

[0003] Existing methods for generating two-dimensional optical tweezers arrays include phase modulation methods based on spatial light modulators (SLMs) and frequency scanning methods based on acousto-optic deflectors (AODs). Among these, two-dimensional orthogonal AOD systems have been widely used in cold atom experimental platforms due to their high refresh rate and ability to rapidly modulate multiple frequency channels.

[0004] However, AOD optical tweezers have the following problems in practical applications:

[0005] Positional nonlinearity: The frequency-angle relationship of AOD is affected by material dispersion and sound velocity non-uniformity, resulting in systematic deviations in the array position;

[0006] Intensity non-uniformity: There are significant differences in diffraction efficiency at different frequencies, resulting in uneven power distribution in the optical tweezers array;

[0007] Manual calibration is inefficient: Traditional methods rely on experimenters to measure point by point and manually fit compensation curves, which is inefficient and difficult to maintain long-term stability.

[0008] Therefore, a calibration method that can be automated and operate in a closed loop is needed to achieve the construction of high-precision, large-scale optical tweezers arrays. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an automatic calibration method for AOD optical tweezers arrays based on neural networks. By integrating a neural network model with a closed-loop control process, the position and intensity calibration of the optical tweezers array is fully automated. Compared with traditional methods that rely on manual point-by-point measurement and manual fitting, this invention requires no manual intervention, significantly reducing operational complexity and time costs, and is particularly suitable for large-scale optical tweezers experimental systems that require frequent calibration.

[0010] The technical solution adopted by this invention to solve its technical problem is: to provide an automatic calibration method for AOD optical tweezers array based on neural networks, comprising the following steps:

[0011] S1. Generate an initial AOD optical tweezer array using an acousto-optic deflector;

[0012] S2, collect an image of the initial AOD optical tweezer array, and pre-process the image;

[0013] S3, input the pre-processed image to a first neural network and a second neural network; wherein the first neural network is used to calculate an average position deviation of the optical tweezer array according to the image; and the second neural network is used to calculate an intensity gain compensation factor of each optical tweezer according to the image;

[0014] S4, calculate a global frequency compensation amount according to the average position deviation and a pre-calibrated frequency-position relationship factor;

[0015] S5, generate an intensity compensation table covering the full frequency domain by an interpolation method based on the intensity gain compensation factor of each optical tweezer;

[0016] S6, update the driving frequency of the AOD according to the global frequency compensation amount, and update the driving amplitude of each channel of the AOD according to the intensity compensation table, and output the corrected AOD optical tweezer array.

[0017] Further, the first neural network used to calculate the average position deviation of the optical tweezer array according to the image comprises the following steps:

[0018] The pre-processed image is subjected to background subtraction and normalization processing;

[0019] The processed image is subjected to one-dimensional projection to obtain row direction projection signal S x (u) and column direction projection signal S y (v), respectively;

[0020] According to the design parameters of the target array, corresponding ideal row position mask M x (u) and ideal column position mask M y (v) are generated;

[0021] The row direction projection signal S x (u), the ideal row position mask M x (u), the column direction projection signal S y (v), and the ideal column position mask M y (v) are spliced into a multi-channel one-dimensional signal;

[0022] The multi-channel one-dimensional signal is input to a pre-trained first neural network, and a set of mapping function coefficients representing spatial distortion is output;

[0023] Based on the set of coefficients, a mapping relationship from ideal design coordinates to actual optical tweezer coordinates is calculated, and a frequency compensation table for compensating spatial distortion is generated in combination with the frequency-position relationship of the AOD.

[0024] Further, the second neural network is configured to calculate an intensity gain compensation factor of each optical tweezer according to the image, including the following steps:

[0025] inputting the preprocessed image as an intensity image into a second neural network, and outputting an intensity distribution model coefficient;

[0026] calculating an intensity compensation factor based on the intensity distribution model coefficient, the intensity compensation factor being determined according to a ratio of a target intensity to a current intensity model calculation value;

[0027] updating a radio frequency driving amplitude of the AOD according to the intensity compensation factor.

[0028] Further, the first neural network is a one-dimensional convolutional neural network or a multi-layer perceptron;

[0029] The mapping function is a quadratic polynomial, and a coefficient set of the mapping function includes α x , β x , γ x , α y , β y , γ y , which are used to describe spatial distortion in the x direction and the y direction.

[0030] Further, the intensity distribution model is a binary quadratic function:

[0031] I'(f x ,f y )=a0+a1f x +a2f y +a3f x 2 +a4f y 2 +a5f x f y , wherein a0 to a5 are coefficients output by the second neural network;

[0032] The second neural network is a convolutional neural network or a fully connected network.

[0033] Further, the correction process is a closed-loop correction, including multiple iterations of frequency calibration and intensity calibration until one of the following exit conditions is met:

[0034] an average position error of the optical tweezer array is lower than a set threshold value and a coefficient of variation of intensity is lower than a set threshold value;

[0035] a preset maximum number of iterations is reached.

[0036] The application also provides an AOD optical tweezer array automatic correction system based on a neural network, including:

[0037] Biorthogonal AOD module, used to generate AOD optical tweezers array;

[0038] An imaging module is used to acquire images of the AOD optical tweezers array;

[0039] The control and processing module is used to execute the processing logic of the frequency calibration and intensity calibration, including running neural network inference, generating frequency compensation tables and intensity compensation factors;

[0040] The radio frequency drive module is used to output a drive signal to the biorthogonal AOD module according to the frequency compensation table and the intensity compensation factor.

[0041] Furthermore, the imaging module includes a high numerical aperture objective lens and a camera, wherein the camera is used to acquire images of the optical tweezers array, and the high numerical aperture objective lens is used to perform imaging magnification on the optical tweezers array.

[0042] Furthermore, the control and processing module integrates a pre-trained first neural network and a second neural network, wherein:

[0043] The first neural network is used to receive multi-channel one-dimensional signals and output the coefficient set of the spatial distortion mapping function;

[0044] The second neural network is used to receive intensity images and output the coefficients of the intensity distribution model;

[0045] The specific steps for calculating the mapping relationship from ideal design coordinates to actual optical tweezer coordinates based on the coefficient set, and generating a frequency compensation table by combining the frequency-position relationship of AOD, include:

[0046] Construct an equally spaced target coordinate grid with the image center as the origin;

[0047] The target coordinate grid is mapped to the actual coordinate grid using the set of mapping function coefficients, wherein the mapping relationship in the x-direction is a combination of linear and quadratic terms, and the mapping relationship in the y-direction is a combination of linear and quadratic terms;

[0048] Based on the linear frequency-position relationship of AOD, the actual coordinate grid is converted into the corresponding driving frequency value;

[0049] A frequency compensation lookup table is generated based on the correspondence between the target coordinates and the converted drive frequency values.

[0050] The present invention has the following beneficial effects:

[0051] (1) The present application realizes the complete automation of the light tweezer array position and intensity correction by integrating the neural network model and the closed-loop control process. Compared with the traditional method which relies on manual point-by-point measurement and manual fitting, the present application does not require human intervention, greatly reduces the operation complexity and time cost, and is particularly suitable for large-scale light tweezer experimental systems which need frequent calibration.

[0052] (2) By real-time analysis of the light tweezer array image through a lightweight neural network, the present application can accurately regress the spatial distortion mapping function and the intensity distribution model coefficient, realize sub-pixel level position correction and highly uniform intensity distribution, and significantly improve the control accuracy and stability of the light tweezer array.

[0053] (3) The lightweight neural network structure adopted by the present application has high inference ability, and the time consumption of single acquisition and processing is short, supporting online real-time correction. The closed-loop iteration mechanism can dynamically adapt to system drift, such as light power fluctuation and AOD efficiency change, to ensure that the light tweezer array maintains the optimal state in long-time experiments.

[0054] (4) The present application uses the strategy of "spatial distortion modeling + linear frequency mapping" to concentrate the nonlinear error in the learnable mapping function, effectively overcoming the systematic deviation caused by factors such as AOD material dispersion, uneven sound speed and optical aberration. The neural network model has strong fault tolerance to image noise and background interference, and is suitable for complex experimental environments.

[0055] (5) The input construction method proposed by the present application, such as one-dimensional projection + ideal mask, avoids the dependence on point-by-point segmentation and centroid fitting, and is suitable for light tweezer arrays of different scales. The system architecture is modular, easy to integrate into existing AOD light tweezer platforms, and has good universality and portability.

[0056] (6) The present application first combines lightweight neural network with AOD light tweezer correction, and proposes a "position-intensity" double closed-loop correction mechanism, breaking through the limitations of traditional static calibration and open-loop compensation. By directly regressing the physical model parameters through the neural network, the present application realizes the interpretability and repeatability of the correction process, and provides reliable technical support for the application of light tweezer technology in quantum computing, cold atom physics and other fields. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any creative labor.

[0058] Figure 1A flowchart of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 1.

[0059] Figure 2 A flowchart of frequency calibration of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 2.

[0060] Figure 3 A flowchart of intensity calibration of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 3.

[0061] Figure 4 A structural diagram of a system of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 4.

[0062] Figure 5 A structural diagram of an initial optical tweezer array of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 5.

[0063] Figure 6 A structural diagram of a frequency-calibrated optical tweezer array of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 6.

[0064] Figure 7 A structural diagram of an intensity-calibrated optical tweezer array of a neural network-based AOD optical tweezer array automatic correction method provided by an embodiment of the present specification is shown in FIG. 7. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0066] In the specification and claims of the present specification and the above-described drawings, the terms "first", "second", "third", etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0067] In the following description, the appearance of terms such as "inner", "outer", "upper", "lower", "left", "right", etc. only indicate the orientation or positional relationship for the convenience of describing the embodiments and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present specification.

[0068] The data involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection of relevant data complies with relevant laws, regulations and standards of relevant countries and regions

[0069] The present application provides a kind of neural network-based AOD light tweezers array automatic correction method and system, to solve the spatial position distortion and intensity distribution uneven problem existing in the generation process of existing AOD light tweezers array.The present application realizes the accurate control and homogenization of light tweezers array by introducing neural network model for frequency calibration and intensity calibration.The specific embodiments of the present application will be described in detail below with reference to the drawings.

[0070] Example 1:

[0071] The present application provides a kind of neural network-based AOD light tweezers array automatic correction method, its core is to realize the accurate correction of AOD light tweezers array spatial position distortion and intensity distribution unevenness through two closed loop processes of frequency calibration and intensity calibration, including the following steps:

[0072] S1, generate initial AOD light tweezers array through acousto-optic deflector;

[0073] S2, collect the image of the initial AOD light tweezers array, and pretreat the image;

[0074] S3, input the pretreated image to the first neural network and the second neural network;Wherein, the first neural network is used to calculate the average position deviation of the light tweezers array according to the image;The second neural network is used to calculate the intensity gain compensation factor of each light tweezers according to the image;

[0075] S4, calculate the global frequency compensation amount according to the average position deviation and the pre-calibrated frequency-position relationship factor;

[0076] S5, generate an intensity compensation table covering the full frequency domain based on the intensity gain compensation factor of each light tweezers through interpolation method;

[0077] S6, update the driving frequency of AOD according to the global frequency compensation amount, and update the driving amplitude of each channel of AOD according to the intensity compensation table, output the corrected AOD light tweezers array.

[0078] 1. Frequency calibration closed loop process

[0079] Frequency calibration aims to correct the spatial positions of the optical tweezer array, making its positions accurate and consistent with the ideal arrangement. The process includes the following steps:

[0080] (a) Generating an initial optical tweezer array and image acquisition: An initial optical tweezer array is generated by the AOD driving device, and the image of the optical tweezer array is acquired by the imaging device (such as a high numerical aperture objective and a camera). This step is the starting point of the correction process, obtaining the actual state of the current optical tweezer array.

[0081] (b) Image preprocessing: The acquired image is preprocessed, including background subtraction and multi-frame averaging. Background subtraction is used to eliminate background noise, and multi-frame averaging can improve the signal-to-noise ratio of the image, providing more accurate data for subsequent analysis.

[0082] (c) One-dimensional projection: The preprocessed image is projected in the row direction and the column direction respectively to obtain the row direction projection signal curve S x (u) and the column direction projection signal curve S y (v). These projection signals reflect the overall distribution of the light spots in each direction, containing spatial distortion information.

[0083]

[0084] (d) Generating ideal position mask: According to the ideal arrangement of the target optical tweezer array, generate the row direction ideal position mask M x (u) and the column direction ideal position mask M y (v) with the same dimension as the projection signal. The ideal position mask is composed of a series of ideal light spot positions represented by a one-dimensional Gaussian function, serving as a reference target for neural network learning.

[0085] (e) Constructing multi-channel one-dimensional input signal: The S x , M x , S y , M y are spliced by channel dimension to construct a multi-channel one-dimensional input signal. This universal input construction method avoids the dependence on point-by-point segmentation and centroid fitting, improving the robustness of the method.

[0086] (f) First neural network inference: The multi-channel one-dimensional input signal is input into the trained first neural network model. The first neural network can be a lightweight one-dimensional convolutional neural network (Conv1D) or a multi-layer perceptron (MLP), whose structure is optimized for efficient processing of one-dimensional signals.

[0087] The first neural network model is a lightweight one-dimensional regression network.

[0088] 3.1 Structure

[0089] Feature extraction: 1D convolution (Conv1D) or light MLP, input is [S x ,M x ,S y ,M y ] (channel dimension can be 2 or 4).

[0090] Parameter head: two regression heads are set respectively to output (α x ,β x ,γ x ) and (α y ,β y ,γ y ).

[0091] Input: one-dimensional sequence length N, channel number 4 (concatenate [S x ,M x ,S y ,M y ] in order)

[0092] Normalization: LayerNorm (4xN)

[0093] Feature extraction (shared trunk, 1D convolution):

[0094] Conv1d (4, 32, kernel=7, padding=3) → RELU → BatchNorm1d (32)

[0095] Conv1d (32, 32, kernel=5, padding=2) → RELU → BatchNorm1d (32) → MaxPool1d (2)

[0096] Conv1d (32, 64, kernel=5, padding=2) → RELU → BatchNorm1d (64) → MaxPool1d (2)

[0097] Conv1d (64, 96, kernel=3, padding=1) → RELU → BatchNorm1d (96) → AdaptiveAvgPool1d (1)

[0098] Flatten: Flatten()

[0099] Desensitization / robust layer: Dropout (p=0.2)

[0100] Fully connected: Linear (96, 64) → GELU → LayerNorm (64)

[0101] Parameter header (x direction): Linear(64, 3) -> output (a x , b x , g x )

[0102] Parameter header (y direction): Linear(64, 3) -> output (a y , b y , g y )

[0103] (g) Output mapping function coefficients: the first neural network model outputs the coefficients of the mapping function used to describe the actual spatial position distortion, including the coefficients a x , b x , g x for the x direction and the coefficients a y , b y , g y for the y direction; wherein the form of the mapping function is: x' = a x x + b x x 2 + g x ; y' = a y y + b y y 2 + g y ; wherein (x, y) is the target coordinate, and (x', y') is the predicted actual coordinate; here, (x, y) is the target coordinate, and (x', y') is the predicted actual coordinate. Substitute x' into the distorted target position, and solve the linear equation set, that is, x can be obtained. This low-order polynomial can effectively depict the spatial distortion and concentrate all nonlinear factors in the spatial mapping.

[0104] (h) Map the target coordinate grid: according to the mapping function, map the target coordinate grid to the predicted actual coordinate grid. This step converts the ideal, distortion-free coordinates into the actual coordinates required for AOD driving.

[0105] In the above spatial reference frame, continue to use the linear frequency-position relationship, but the input is the distorted x':

[0106]

[0107] wherein

[0108] k x ,k y are constant proportional coefficients (obtained by one-time calibration, which can be given by system geometry and AOD physical constants or obtained by experimental fitting).

[0109] Because of zero-point alignment (center frequency = image center), constant terms are automatically absorbed into the γ and zero-point definitions, eliminating the need for additional frequency intercepts.

[0110] The core idea is to leave all nonlinearity in the "spatial mapping" while keeping the mapping from frequency to the "distorted space" linear; this makes it robust, easy to interpret, and easy to maintain.

[0111] (i) Generating a Frequency Compensation Lookup Table (LUT): Based on the linear frequency-position relationship, the required driving frequency is calculated according to the predicted actual coordinates, and a frequency compensation lookup table (LUT) is generated. The linear frequency-position relationship is: f x,i =f 0x +k x *x i ';f y,j =f 0y +k y *y j '; where f 0x ,f 0y k is the center frequency of AOD. x ,k y It is a frequency-position scaling factor; by maintaining the linearity of the frequency mapping in the distorted spatial reference frame, it improves the robustness, interpretability, and maintainability of the system.

[0112] (j) Update AOD drive frequency: Update the output frequency of the AOD drive device according to the frequency compensation LUT to complete the frequency calibration. This step sends the calibrated frequency information to the AOD to adjust the actual position of the optical tweezers.

[0113] 2. Strength calibration closed-loop process

[0114] Intensity calibration aims to adjust the intensity distribution of the optical tweezers array to achieve uniform intensity. The intensity calibration closed-loop process is initiated after position calibration is completed and includes the following steps:

[0115] (k) Acquire intensity image of the optical tweezers array after frequency update: After frequency calibration is completed, acquire intensity image of the optical tweezers array. At this point, the position of the optical tweezers has been basically corrected, and the focus is on the intensity distribution.

[0116] (l) Second Neural Network Inference: The intensity image is input into a trained second neural network model. This second neural network can be a convolutional neural network (CNN) or a fully connected network (FCN) to analyze the intensity distribution characteristics of the optical tweezers array.

[0117] The second neural network model is as follows:

[0118] Input: Array intensity distribution I ij ′+corresponding frequency coordinate (f x ,fy );

[0119] Network structure: light neural network (MLP or CNN);

[0120] Output: six coefficients (a0, a1, a2, a3, a4, a5).

[0121] Input: two-dimensional array of optical tweezer images (background subtraction, normalization), fixed size N x N.

[0122] Feature extraction:

[0123] Conv2d(1,32,kernel=5,padding=2)→ReLU→BatchNorm2d(32)

[0124] Conv2d(32,64,kernel=3,padding=1)→ReLU→BatchNorm2d(64)→MaxPool2d(2)

[0125] Conv2d(64,128,kernel=3,padding=1)→ReLU→BatchNorm2d(128)→MaxPool2d(2)

[0126] Flatten: Flatten()→get length vector

[0127] Fully connected layer:

[0128] Linear(128·(N / 4)·(N / 4),256)→ReLU→Dropout(0.3)

[0129] Linear(256,64)→ReLU

[0130] Output layer: Linear(64,6)→output binary quadratic coefficients (a0, a1, a2, a3, a4, a5).

[0131] (m) Output intensity distribution model coefficients: the second neural network model outputs the coefficients a0, a1, a2, a3, a4, a5 of the binary quadratic function describing the intensity-frequency distribution law; wherein the form of the binary quadratic function is: I'(f x ,f y )=a0+a1*f x +a2*f y +a3*f x 2 +a4*f y 2 +a5*f x *fy ; where I'(f x ,f y ) is the predicted optical tweezer intensity at frequency (f x ,f y ); this model is used to quantify the intensity non-uniformity caused by the frequency dependence of AOD diffraction efficiency, optical system aberration, and cross-AOD frequency coupling, etc.

[0132] (n) Calculate amplitude compensation factor: according to the target intensity I target and the predicted intensity I'(f x ,f y ), calculate the amplitude compensation factor G(f x ,f y ) of the whole domain: G(f x ,f y )=sqrt(I target / I'(f x ,f y )).

[0133]

[0134] The compensation factor reflects the amplitude ratio required for each optical tweezer to achieve the target intensity.

[0135] Update the radio frequency drive amplitude according to the intensity compensation factor. A old (f x ,f y ) is the initial radio frequency drive amplitude.

[0136]

[0137]

[0138]

[0139] (o) Generate amplitude compensation lookup table and update AOD drive amplitude: generate amplitude compensation lookup table according to the amplitude compensation factor G(f X ,f y ), and update the output amplitude of the AOD drive device. This step will issue the calculated amplitude compensation information to the AOD to adjust the intensity of the optical tweezers.

[0140] (p) Repeat iteration until the condition is met: repeat steps (k) to (o) until the intensity uniformity of the optical tweezer array meets the preset condition. The exit condition can include that the intensity coefficient of variation within the array is lower than the preset threshold (such as 5%-10%), or the maximum iteration number (such as 5 times) is reached. The whole correction process is a closed-loop correction, which ensures that the correction effect meets the requirements through multiple iterations.

[0141] Example 2: Neural network-based AOD optical tweezer position and intensity closed-loop calibration

[0142] To verify the effectiveness of the method of the present application, a neural network-based AOD optical tweezer position and intensity closed-loop calibration experiment was conducted. This embodiment describes in detail the experimental device, experimental conditions and calibration results.

[0143] 1. Experimental device and conditions

[0144] Hardware configuration: The experimental device mainly includes a double-quadrature AOD, a DDS radio frequency driver, a high numerical aperture (NA) objective lens, an sCMOS camera (resolution 2048x2048 pixels, calibrated pixel to micrometer (px->µm) conversion matrix), and a control PC for running neural network inference and waveform issuance.

[0145] Experimental conditions: The laser wavelength used in the experiment is 813 nm, and the optical tweezer array size constructed is 10x10.

[0146] 2. Position calibration process and results

[0147] Initial array generation: By setting the frequency points of the AOD, a 10x10 optical tweezer array is generated. In the initial state, it is observed that the optical tweezers at the edges and corners of the array have obvious position deviation, and the overall presents a curved distortion.

[0148] Image acquisition and input construction: The sCMOS camera acquires multiple frames of optical tweezer array images and performs background subtraction and multi-frame averaging preprocessing. Subsequently, one-dimensional projection is performed on the preprocessed images along the row and column directions to obtain one-dimensional signals S x and S y . Combined with the pre-generated ideal masks M x and M y , the multi-channel one-dimensional input signal is constructed.

[0149] Neural network inference and frequency mapping update: The constructed multi-channel one-dimensional input signal is input into the pre-trained first neural network model. The network outputs the quadratic function coefficients describing the spatial distortion. According to these coefficients, the actual coordinates x' and y' are calculated, and the frequency linear relationship is maintained, the frequency mapping table is updated and issued to the DDS radio frequency driver.

[0150] Calibration results: After one closed-loop calibration, the position deviation of the optical tweezer array is significantly reduced to sub-pixel level, and the overall arrangement tends to be regular. Experimental data shows that the average position deviation is reduced from about 3.3µm to about 0.1µm, indicating that the position calibration effect is significant.

[0151] 3. Intensity calibration process and results

[0152] Initial intensity distribution: In the array after position correction, the intensity of the optical tweezers still exhibits a non-uniformity, with the center being stronger and the edges being weaker.

[0153] Neural network inference and compensation LUT generation: Intensity images of the optical tweezers array after position correction are acquired and input into a pre-trained second neural network model. This network outputs the coefficients of a binary quadratic function (I'(f)) describing the intensity distortion. x ,f y )=a0+a1*f x +a2*f y +a3*f x 2 +a4*f y 2 +a5*f x *f y The compensation factor is calculated based on these coefficients, and an amplitude compensation lookup table is generated to update the output amplitude of the DDS RF driver.

[0154] Closed-loop update and calibration results: Reacquire array images and evaluate the coefficient of variation (CV). If the CV is greater than a preset threshold (e.g., 10%), continue iteration; otherwise, exit the closed loop. After two rounds of iteration, the array intensity uniformity is significantly improved, and the coefficient of variation decreases from 27% to 6%.

[0155] 4. Summary of Results

[0156] This invention achieves high-precision position and intensity calibration of the AOD optical tweezers array through closed-loop correction:

[0157] Position correction: The average deviation was reduced from approximately 3.3 µm to approximately 0.1 µm.

[0158] Strength correction: The coefficient of variation decreased from 27% to 6%.

[0159] Closed-loop efficiency: The time for a single data acquisition and neural network inference is less than 200ms, which can meet the requirements of real-time experiments.

[0160] These results fully demonstrate the effectiveness and practicality of the present invention in the automatic calibration of AOD optical tweezer arrays.

[0161] Example 3:

[0162] This invention also provides a system for implementing the above-described automatic calibration method for AOD optical tweezers arrays. This system, through a modular design, integrates functions such as optical tweezers generation, image acquisition, data processing, and calibration signal output, achieving automated and high-precision calibration of the optical tweezers array. The system mainly includes the following modules:

[0163] 1. Biorthogonal AOD module

[0164] Function: Used to receive the driving signal output by the radio frequency driving module, and generate AOD optical tweezer array through acousto-optic deflection effect. This module is the physical generation unit of the optical tweezer array, and can accurately control the deflection angle and diffraction efficiency of the light beam according to the input frequency and amplitude signal, thereby forming the required optical tweezer array.

[0165] 2. Imaging module

[0166] Composition: Includes a high numerical aperture (NA) objective lens and a camera (such as an sCMOS camera).

[0167] Function: Used to collect the image of the AOD optical tweezer array. The high-NA objective lens is used to image and magnify the optical tweezer array, ensuring that the image has sufficient resolution and clarity for subsequent image processing and analysis. The camera is responsible for converting the optical image into a digital signal for use by the control and processing module. This module can collect the original image and intensity image of the optical tweezer array, providing data input for frequency calibration and intensity calibration.

[0168] 3. Control and processing module

[0169] Composition: Integrates a pre-trained first neural network and a second neural network, and is usually implemented by a control PC.

[0170] Function: As the core of the system, it is used to execute the processing logic of frequency calibration and intensity calibration. Specific functions include:

[0171] Run neural network inference: The first neural network is used to receive multi-channel one-dimensional signals (generated by image preprocessing and projection), and output a set of coefficients (α x , β x , γ x , α y , β y , γ y ) of the mapping function representing spatial distortion. The second neural network is used to receive intensity images and output coefficients (a0 to a5) of the intensity distribution model.

[0172] Generate frequency compensation table: Based on the coefficient set output by the first neural network, calculate the mapping relationship from the ideal design coordinates to the actual optical tweezer coordinates, and generate the frequency compensation lookup table (LUT) used to compensate for spatial distortion in combination with the frequency-position relationship of the AOD. The specific steps include constructing an equidistant target coordinate grid, mapping the target coordinates to actual coordinates using the mapping function, converting the actual coordinates to driving frequency values based on the linear frequency-position relationship of the AOD, and finally generating the frequency compensation lookup table.

[0173] Generate intensity compensation factor: Based on the intensity distribution model coefficients output by the second neural network, calculate the intensity compensation factor G(f X ,f y), and generate an amplitude compensation lookup table.

[0174] Downlink driving signal: convert the generated frequency compensation table and amplitude compensation lookup table into the driving signal required by the AOD driving device, and send it to the radio frequency driving module.

[0175] 4. Radio frequency driving module

[0176] Composition: usually includes DDS (direct digital frequency synthesizer) or AWG (arbitrary waveform generator).

[0177] Function: according to the frequency compensation table and intensity compensation factor issued by the control and processing module, output the driving signal to the double quadrature AOD module. This module is responsible for converting the digitized correction information into analog radio frequency signal, accurately controlling the driving frequency and driving amplitude of AOD, so as to realize the position correction and intensity uniformization of optical tweezer array.

[0178] Through the cooperative work of the above-mentioned modules, the system of the present application can realize the automatic, high-precision closed-loop correction of AOD optical tweezer array, and significantly improve the performance and application range of optical tweezer array.

[0179] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, as long as it is within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made shall be included in the protection scope of the present application.

Claims

1. An automatic calibration method for AOD optical tweezers array based on neural networks, characterized in that, Includes the following steps: S1. Generate an initial AOD optical tweezer array using an acousto-optic deflector; S2. Acquire images of the initial AOD optical tweezer array and preprocess the images; S3. Input the preprocessed image into the first neural network and the second neural network; wherein, the first neural network is used to calculate the average positional deviation of the optical tweezers array based on the image; the second neural network is used to calculate the intensity gain compensation factor of each optical tweezer based on the image; S4. Calculate the global frequency compensation table based on the average position deviation and the pre-calibrated frequency-position relationship factor, specifically including: The preprocessed image is then subjected to background subtraction and normalization. The processed image is subjected to one-dimensional projection to obtain the row direction projection signal S. x (u) and column direction projection signal S y (v); Based on the design parameters of the target array, generate the corresponding ideal row position mask M. x (u) and ideal column position mask M y (v); Projecting signal S in the direction of travel x (u), Ideal row position mask M x (u), column direction projection signal S y (v) Ideal column position mask M y (v) splicing into a multi-channel one-dimensional signal; A multi-channel one-dimensional signal is input into a pre-trained first neural network, which outputs a set of mapping function coefficients representing spatial distortion. Based on the coefficient set, the mapping relationship from the ideal design coordinates to the actual optical tweezers coordinates is calculated, and combined with the frequency-position relationship of AOD, a global frequency compensation table for compensating for spatial distortion is generated. S5. Based on the intensity gain compensation factor of each optical tweezer, an intensity compensation table covering the entire frequency domain is generated by interpolation. S6. Update the driving frequency of AOD according to the global frequency compensation table, and update the driving amplitude of each channel of AOD according to the intensity compensation table, and output the corrected AOD optical tweezer array.

2. The method for automatic calibration of AOD optical tweezers array based on neural network according to claim 1, characterized in that, S5. Based on the intensity gain compensation factor of each optical tweezer, an intensity compensation table covering the entire frequency domain is generated using an interpolation method, including the following steps: The preprocessed image is input as an intensity image into the second neural network, which outputs intensity distribution model coefficients. The intensity compensation factor is calculated based on the intensity distribution model coefficients, and the intensity compensation factor is determined according to the ratio of the target intensity to the current intensity model calculated value; The AOD's radio frequency drive amplitude is updated based on the intensity compensation factor to generate an intensity compensation table covering the entire frequency domain.

3. The method for automatic calibration of AOD optical tweezers array based on neural network according to claim 1, characterized in that, The first neural network is a one-dimensional convolutional neural network or a multilayer perceptron; The mapping function is a quadratic polynomial, and its coefficient set includes α. x β x γ x α y β y γ y , used to describe spatial distortion in the x and y directions.

4. The method for automatic calibration of AOD optical tweezers array based on neural network according to claim 2, characterized in that, The intensity distribution model is a bivariate quadratic function: I'(f x ,f y )=a0+a1f x +a2f y +a3f x 2 +a4f y 2 +a5f x f y , where a0 to a5 are the coefficients of the second neural network output; The second neural network is a convolutional neural network or a fully connected network.

5. The method for automatic calibration of AOD optical tweezers array based on neural network according to claim 1, characterized in that, The correction process is a closed-loop correction, including multiple iterations of frequency and intensity calibration until one of the following exit conditions is met: The average positional error of the optical tweezers array is lower than a set threshold and the intensity variation coefficient is lower than a set threshold. The preset maximum number of iterations has been reached.

6. An automatic calibration system for an AOD optical tweezers array for implementing the method as described in any one of claims 1-5, characterized in that, include: Biorthogonal AOD module, used to generate AOD optical tweezers array; An imaging module is used to acquire images of the AOD optical tweezers array; The control and processing module is used to execute the processing logic of the frequency calibration and intensity calibration, including running neural network inference, generating frequency compensation tables and intensity compensation factors; The radio frequency drive module is used to output a drive signal to the biorthogonal AOD module according to the frequency compensation table and the intensity compensation factor.

7. The system according to claim 6, characterized in that, The imaging module includes an objective lens and a camera. The camera is used to acquire images of the optical tweezers array, and the objective lens is used to magnify the optical tweezers array.

8. The system according to claim 6, characterized in that, The control and processing module integrates a pre-trained first neural network and a second neural network, wherein: The first neural network is used to receive multi-channel one-dimensional signals and output the coefficient set of the spatial distortion mapping function; The second neural network is used to receive intensity images and output the coefficients of the intensity distribution model; The specific steps for generating a global frequency compensation table to compensate for spatial distortion include: Calculating the mapping relationship from ideal design coordinates to actual optical tweezer coordinates based on the coefficient set, and combining this with the frequency-position relationship of the AOD. Construct an equally spaced target coordinate grid with the image center as the origin; The target coordinate grid is mapped to the actual coordinate grid using the set of mapping function coefficients, where the mapping relationship in the x-direction is a combination of linear and quadratic terms, and the mapping relationship in the y-direction is a combination of linear and quadratic terms; Based on the linear frequency-position relationship of AOD, the actual coordinate grid is converted into the corresponding driving frequency value; A global frequency compensation table is generated based on the correspondence between the target coordinates and the converted drive frequency values.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 5.

Images