Spatial multiplexing acoustic hologram design method based on physical model neural network

Abstract

The invention discloses a spatial multiplexing acoustic hologram design method based on a physical model neural network. Parameters are determined according to a target spatial multiplexing design scene, a sound wave propagation model and a deep learning model are established, the input of the deep learning model is a random noise image, the output of the deep learning model is a transducer array selection area graph and a phase plane graph, and the sound wave propagation model and the deep learning model are combined to establish an overall model to be combined with a target sound field graph for training; and after training, obtaining a transducer array selection area graph and a phase plane graph. According to the method, the problem that artifacts exist in multichannel information due to lack of global optimization in a traditional angular spectrum iterative design method is solved; the method overcomes the defects that an existing deep learning method consumes time and labor and is poor in robustness, and specific parameters cannot be designed according to the design requirements of the target holographic sound field when a data set is constructed, optimization design can be carried out on multiple target holographic sound fields at the same time, and any holographic sound field can be designed in a self-adaptive mode.

Description

Technical Field

[0001] The invention relates to the technical field of acoustic holography, and in particular to a method for designing a spatial multiplexing acoustic hologram based on a physical model neural network. Background Art

[0002] Acoustic holography is an extension and development of computational holography in the field of acoustics. It records the acoustic field information through modulation functional units arranged in two-dimensional space, and reconstructs the acoustic information with high fidelity under the excitation of a specific incident acoustic field. Therefore, it can be used for high-degree-of-freedom modulation of any complex acoustic field. Compared with the traditional phased array system, the passive acoustic holographic device reconstructs the acoustic field in three dimensions through millions of discrete modulation functional units arranged in two-dimensional space, so it has higher resolution and is currently widely used in applications such as acoustic communication, acoustic-assisted particle manipulation and tissue regeneration. However, conventional acoustic holographic devices prepared by 3D printing can only achieve specific acoustic responses once prepared. This feature greatly limits the modulation freedom and application scope. Based on this, existing research has improved the modulation freedom of the holographic acoustic field by frequency multiplexing, spatial multiplexing, and spatiotemporal multiplexing. However, how to reduce the superposition of artifacts between multiple holographic acoustic fields while ensuring high-resolution reconstruction of the target holographic acoustic field is an urgent problem to be solved in the field of multiplexed acoustic holography.

[0003] Most of the existing algorithms are based on the classical iterative angular spectrum method, which has shown excellent design and reconstruction effects in the design of a single acoustic hologram. However, when designing an acoustic phase hologram based on multiple target holographic sound fields at the same time, it is often easy to cause large artifacts due to the lack of a global optimization method. In addition, due to design limitations, most of the existing spatial multiplexing technologies pre-set the situation of the incident sound field during design, and do not perform collaborative optimization of the incident sound wave situation.

[0004] In recent years, deep learning has been gradually applied to the design of acoustic holographic devices. By learning the rules of the data set, the end-to-end mapping of the target holographic surface and the design surface can be obtained. The corresponding mapping relationship between multiple target holographic sound fields and the reuse of functional units can be realized, which has the advantages of high efficiency and high precision. However, the above method is data-driven, and the performance of the neural network depends on the diversity and accuracy of the training data set. A large data set is often required for complex acoustic holographic designs, and building a large data set requires a lot of time and effort. At the same time, deep learning has poor generalization and poor applicability to flexible and changeable acoustic holographic designs. Summary of the invention

[0005] In order to solve the problems existing in the background technology, the present invention provides a design method based on physical model deep learning for spatial multiplexing acoustic holograms.

[0006] The present invention is used to solve the problem of artifacts in multi-channel information due to the lack of global optimization in the traditional angular spectrum iterative design method; as well as the shortcomings of existing deep learning methods such as being time-consuming and labor-intensive when constructing data sets, having poor robustness, and being unable to design specific parameters according to the design requirements of the target holographic sound field.

[0007] To achieve the above object, the technical solution provided by the present invention is as follows:

[0008] Step 1: Determine various parameters and the distance and medium of sound wave propagation according to the target spatial reuse design scenario and design requirements;

[0009] Step 2: Establishing a sound wave propagation model, wherein the sound wave propagation model is used to transfer the sound field image of one plane to another plane through the device of the actual scene under the target space multiplexing design scenario of specific requirements, thereby obtaining the sound field image of the other plane;

[0010] The sound wave propagation model is specifically used to transfer the image output by the deep learning model to the target plane through the actual scene, and then obtain the sound field image of the target plane.

[0011] The spatial multiplexing acoustic holographic design targeted by the present invention is a near-field acoustic problem. The sound wave propagation model is based on the angular spectrum theory in Fourier acoustics, which extrapolates the sound field from one plane to another.

[0012] Step 3: Build a deep learning model. The input of the model is a fixed random noise image, and the output is a transducer array selection map and a phase plane map. Constrain the output image to make it meet the parameters preset in step 1.

[0013] Step 4: Combine the sound wave propagation model and the deep learning model to build an overall model and train it with the previously known target sound field map. After training, the transducer array selection map and phase plane map are obtained for presenting the target sound field map through the device in the target space multiplexing design scenario.

[0014] In a specific implementation, the control transducer array is switched according to the transducer array selection diagram, and the phase holographic plate is arranged according to the phase plane diagram, so that the transducer array can present the target sound field through the phase plate after emitting ultrasonic waves.

[0015] After training, the deep learning model can achieve the optimal transducer array excitation area design and phase distribution design on the acoustic phase hologram plate for a specific multi-target holographic sound field, and can display the most complete spatially multiplexed acoustic hologram with the least artifacts.

[0016] The target space multiplexing scenario refers to the rapid reconstruction of the spatial sound field for scenarios requiring acoustic manipulation, such as thermal ablation and neural stimulation under transcutaneous acoustic thermal effect.

[0017] The transducer array is composed of multiple ultrasonic transducers tightly arranged in an array on the same plane. Specifically, the probe planes of multiple ultrasonic transducers are tightly arranged in an array on the same plane, and each ultrasonic transducer has only two states, on and off, and is a binary transducer. The transducer array selection map is a distribution map of the on or off state of each ultrasonic transducer.

[0018] The various parameters of step 1 specifically include the spatial excitation area of ​​the transducer array, the length and width of the phase hologram, the production resolution limit of the phase hologram, the image matrix size of the model output, the sound wave propagation distance, frequency, and propagation medium of the transducer array, etc.

[0019] The sound wave propagation model of step 2 is specifically:

[0020] Step 2.1, set the transducer array, the phase hologram plate and the target plane to be spaced in sequence along the positive direction of the z-axis depth direction. The sound pressure field p1 emitted by the transducer array propagates in the positive direction along the z-axis depth direction to the sound pressure field p2 of the phase hologram plate located on the z=D1 plane is:

[0021]

[0022] Among them, x and y represent the position variables in the spatial Cartesian coordinate system, k x , k y represent the variables in the wave number space, represents the wave field extrapolation function of Fourier acoustics, p2(x, y, D1) represents the sound pressure field on the z=D1 plane in the spatial Cartesian coordinate system, and k represents the wave number of the sound wave in the sound propagation medium with a wavelength of λ;

[0023] Step 2.2, the sound field p2 transmitted from the transducer array to the phase hologram undergoes target phase modulation after passing through the phase hologram, so that the transmitted sound field of the phase hologram introduces the phase distribution The sound pressure field p2′ after passing through the phase holographic plate is obtained according to the following formula:

[0024]

[0025] Among them, e represents a natural constant, and i represents an imaginary unit;

[0026] Step 2.3, the modulated sound pressure field p2′ continues to propagate forward along the z-axis depth direction to the target plane located on the z=D2 plane, and the sound pressure field p3 on the target plane is obtained by processing according to the following formula:

[0027]

[0028] in, is the wave field extrapolation function of Fourier acoustics.

[0029] The sound wave propagation model is used to calculate the sound field distribution when the sound field is transferred from one surface to another surface.

[0030] The deep learning model adopts a U-net++ network structure, and performs specific constraint processing on the output image of the U-net++ network structure according to the design of the acoustic holographic device.

[0031] In step 3, the deep learning model outputs multiple transducer array selection maps / phase plane maps, that is, both the transducer array selection map and the phase plane map can be multiple or one, which are obtained by adjusting the number of output channels in the U-net++ network structure.

[0032] In step 3, the specific steps are:

[0033] Step 3.1, using a randomly generated Gaussian noise image of size 512×512 as a random noise image, first use the U-net++ network structure to process the random noise image, and the U-net++ network structure outputs the encoding amplitude distribution A1 corresponding to the transducer array as the transducer array selection map and the phase modulation distribution corresponding to the phase hologram As a phase plane diagram;

[0034] If the design goal is multiple target holographic sound fields, such as Figure 4 As shown, the number of output channels can be increased as required. For example, if three target holographic sound fields need to be designed, the output will be three sound field amplitude distributions A1, A2, A3 and phase change distributions. In summary, it is an image of size 512×512 with 4 channels.

[0035] Step 3.2, then perform an average pooling operation on the transducer array selection image A1, convert the image into 8×8 size, then perform nearest neighbor upsampling to expand the image to 512×512, and finally perform binarization processing; at the same time, perform low-pass filtering on the phase distribution image to filter out the high-frequency components, so as to meet the accuracy limit requirements for the production of the phase holographic plate.

[0036] The training in step 4 is specifically:

[0037] 4.1. First, the transducer array selection map and phase plane map output by the deep learning model are used as the sound field map of the plane in the actual scene of the target spatial multiplexing design scene, and are input into the sound wave propagation model to obtain the holographic sound field map after the actual scene of the target spatial multiplexing design scene through calculation;

[0038] 4.2. Subtract the holographic sound field image from the target sound field image to obtain the loss;

[0039] 4.3. The loss is then input into the deep learning model for back propagation optimization. The optimized deep learning model outputs the transducer array selection map and phase plane map again, and then returns to step 4.1 for loop iteration;

[0040] 4.4. Repeat steps 4.1 to 4.3 for multiple times until the loss is less than / reaches the preset loss threshold, and take the transducer array selection map and phase plane map obtained by the final optimization as the output result.

[0041] In step 4, the loss uses mean square error (MSE) as the loss function of the network and measures the gap between the hologram optimized by the network and the target hologram. The formula is as follows:

[0042]

[0043] in, is the amplitude part of the holographic sound field image derived from the sound wave propagation model, A is the amplitude part of the target holographic sound field, abs() represents the absolute value function, Represents the holographic sound field envelope distribution derived from the sound wave propagation model.

[0044] The present invention takes multiple holographic sound fields as optimization targets, and designs the excitation matrix of the binary partitioned transducer and the fused phase holographic plate simultaneously to meet the realization of multiple target holographic sound fields under multiplexing conditions; after the design is completed, different incident sound fields are excited in space by changing the binary transducer, and after working together with the fused phase holographic plate, switching between multiple preset holographic sound fields can be realized on the target image plane.

[0045] The present invention combines the Fourier acoustic model with the deep learning model, builds an overall model and completes the training, thereby optimizing the design of multiple target holographic sound fields at the same time, and linking the amplitude and phase distribution of the deep learning design with the target hologram, so that it can adaptively design any holographic sound field.

[0046] Compared with the existing design method, the present invention has the following advantages:

[0047] The network model used in the present invention is unsupervised learning, and the goal is to enable the network to learn to generate the best amplitude distribution and phase distribution for the target hologram. The addition of the sound wave propagation model allows the image generated by the network to follow the physical laws of sound wave propagation and learn to generate holographic images, so there is no need to build a data set that is time-consuming and labor-intensive. For each design, it is a new training process, which can adapt to any acoustic holographic design and has better robustness.

[0048] The present invention generates a target optimized image and sets certain constraints on the generated image to meet most design requirements. Only one forward propagation of the sound wave is required to transmit the sound wave from the transducer plane, through the phase hologram plate and forwardly propagate to the target plane, without the need for multiple forward-backward propagations as in the traditional iterative angular spectrum method. At the same time, the optimization method can generate phase distribution and amplitude distribution at the same time, which makes up for the defects of the traditional iterative angular spectrum method. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A logical process diagram of the method of the present invention;

[0050] Figure 2 It is a schematic diagram of the actual scene of the present invention;

[0051] Figure 3 It is an overall illustration of the method of the present invention;

[0052] Figure 4 A diagram of the spatial multiplexing acoustic holography technology based on a transducer array and a phase holographic plate implemented by the present invention. Figure 5 To randomly generate an amplitude distribution map of a transducer array; Figure 6 This is a diagram of the phase and amplitude distribution results of the present invention and the iterative angular spectrum method. DETAILED DESCRIPTION

[0053] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0054] like Figure 1 As shown, in the specific implementation of the present invention, an ultrasonic transducer array is used as an ultrasonic excitation device. Different from the traditional ultrasonic transducer device, the transducer can independently control the excitation of each transducer unit through a switching circuit, thereby achieving control of the amplitude of the incident sound field.

[0055] Specifically, the transducer array is composed of multiple ultrasonic transducers tightly arranged in the same plane. Specifically, the probe planes of multiple ultrasonic transducers are tightly arranged in the same plane, and each ultrasonic transducer has only two states, on and off, and is a binary transducer. The transducer array selection map is a distribution map of the on or off state of each ultrasonic transducer.

[0056] In addition, the phase holographic plate is a 3D-printed acoustic metasurface that controls the phase of the transmitted sound wave by spatially arranging the thickness of each height unit.

[0057] The method combines the acoustic wave propagation model with the neural network. Based on the propagation between surfaces, deep learning can learn the laws of data. By combining the acoustic wave propagation model with deep learning, deep learning can be used to simultaneously optimize the design of the excitation matrix of the transducer array and the phase distribution of the phase hologram.

[0058] Embodiments of the present invention are as follows:

[0059] Example 1

[0060] Step 1, determine the parameters of the transducer array and the phase holographic plate as well as the distance and medium of sound wave propagation.

[0061] like Figure 2 As shown, the emitting surface and phase plate of the transducer array are both squares with a side length of L, where L = 50 mm; according to the manufacturing accuracy limit of the phase holographic plate (200 μm) and the image matrix size requirements of U-net++, the pixel size of the image is set to dx = (50 / 512) mm; the sound wave is excited from the ultrasonic transducer at the z = 0 mm plane, and is transmitted along the positive direction of the z axis to the phase plate at z = D1 = 20 mm for phase modulation, and then continues to propagate along the positive direction of the z axis to the target holographic plane at z = D2 = 40 mm; in particular, the ultrasonic frequency f0 = 2.2 MHz excited by the transducer array, the propagation medium of the sound wave is water, and the sound speed c = 1500 m / s.

[0062] Step 2: Determine the acoustic propagation model.

[0063] Step 2.1, setting the transducer array, the phase holographic plate and the target plane to be spaced in sequence along the positive direction of the z-axis depth direction, and the dimensions of the three structures in the x and y directions are the same, wherein the transducer array is located in the z=0 plane;

[0064] Assuming that the sound pressure of the sound wave excited by the transducer array is A1(x, y), the frequency is f, the phase is 0, and the transducer array plane (z=0) is the initial plane, then the sound pressure field p1 on the transducer array plane can be expressed as:

[0065] p1(x,y,z=0)=A1(x,y)

[0066] Based on Fourier acoustic theory, the angular spectrum of the sound pressure field p1 in the z=0 plane can be calculated:

[0067]

[0068] The sound pressure field p1 emitted by the transducer array propagates forward along the z-axis depth direction to the sound pressure field p2 of the phase hologram plate located on the z=D1 plane:

[0069]

[0070] Step 2.2, the sound field p2 transmitted from the transducer array to the phase hologram plate undergoes target phase modulation after passing through the phase hologram plate, that is, the sound field will undergo a phase mutation after passing through the phase hologram plate, and the phase distribution on the phase hologram plate is Introducing phase distribution into the transmitted sound field of the phase holographic plate The sound pressure field p2′ after passing through the phase holographic plate is obtained according to the following formula:

[0071]

[0072] Step 2.3, the modulated sound pressure field p2′ continues to propagate forward along the z-axis depth direction to the target plane located on the z=D2+D1 plane, and the sound pressure field p3 on the target plane is obtained by processing according to the following formula:

[0073]

[0074] in, is the wave field extrapolation function of Fourier acoustics.

[0075] Based on the size determined in step 1 and Figure 2 The direction of the lower left corner establishes a spatial rectangular coordinate system, and the coordinates are specifically set to L represents the side length of the emitting surface of the transducer array and the phase plate. To transform the spatial domain coordinates into the frequency domain, it is also necessary to establish the frequency domain coordinates. At this time, set the left side of the frequency to

[0076] Step 3: Build a deep learning model.

[0077] Step 3.1, use the pytorch framework to build a U-net++ model. According to the design requirements, the network input is 1 channel and the output is 4 channels. The depth of the U-net++ model used is 5 layers, and the number of channels in each layer in downsampling is [64, 128, 256, 512, 1024]. Leakyrelu() is used as the activation function, and BatchNorm2d() is used as the normalization processing model of the network.

[0078] The network training uses ADAM as the optimizer, and the learning rate decay uses cosine annealing decay. The input image is a two-dimensional Gaussian distribution noise image processed by the torch.randn() function, and the output is 4 channels.

[0079] Step 3.2: Process the output image according to design requirements.

[0080] In the specific implementation, an 8×8 transducer array is used to output a 521×521 sound wave amplitude distribution image A. iAn average pooling operation is performed to convert the image into an 8×8 size to match the network output matrix and the excitation matrix size of the transducer array. At the same time, when performing angular spectrum method extrapolation, the 8×8 transducer excitation matrix is ​​expanded to a size of 512×512 through nearest neighbor upsampling, so that the calculation process conforms to the 521×521 size during matrix calculation. At the same time, the excitation matrix of the transducer array is limited to only 8×8 matrix variables.

[0081] In addition, the back end of the transducer array in the present invention is a switch circuit, so there are only two states of "on" and "off", so the 8×8 excitation matrix is ​​binarized. At the same time, due to the certain precision limit of the phase holographic plate production (generally 200μm), the phase distribution image on the output phase holographic plate needs to be filtered during the optimization process to filter out the high-frequency detail components.

[0082] Among the 4 channels, Figure 4 As shown, the first three channels correspond to the amplitude regulation, and the output three transducer array selection images corresponding to the three target holographic sound fields are mainly the excitation matrix of the transducer array. Therefore, it is necessary to perform average pooling processing on the output images of the first three channels, and use the torch.AvgPool2d(kernel_size=64, stride=64) function to convert the original 512×512 image into an 8×8 image, and then perform binarization processing to convert it into an 8×8 encoding excitation matrix of the ultrasonic transducer array.

[0083] The fourth channel corresponds to the phase distribution of the phase hologram. It outputs a fused phase plane map and filters the phase image. The image is processed by torch.fft.fft2() and torch.fft.fftshift() in turn. At this time, the low-frequency part is located in the central area of ​​the image, so a circular mask with a certain radius is used to remove the external frequency signal to achieve low-pass filtering. The radius used here is 48. The filtered image is smoother and more concentrated, which is convenient for the subsequent production of the phase plate.

[0084] Step 4: Combine the deep learning model with the sound wave propagation model.

[0085] 4.1. First, the transducer array selection map and phase plane map output by the deep learning model are used as the sound field map of the plane in the actual scene of the target space multiplexing design scene, and are input into the sound wave propagation model, and the holographic sound field map after the actual scene of the target space multiplexing design scene is obtained is output;

[0086] 4.2. Then, the holographic sound field image and the target sound field image are subtracted to obtain the loss; the output image of the neural network can be calculated through steps 1 and 2 to obtain the reconstructed holographic sound field image on the target plane. The loss uses mean square error (MSE) as the loss function of the network and measures the gap between the hologram optimized by the network and the target hologram. The formula is as follows:

[0087]

[0088] 4.3. The loss is then input into the deep learning model for back propagation optimization. The optimized deep learning model outputs the transducer array selection map and phase plane map again, and then returns to step 4.1 for loop iteration;

[0089] 4.4. Repeat steps 4.1 to 4.3 for multiple times until the loss is less than / reaches the preset loss threshold, and use the transducer array selection map and phase plane map obtained last time as the output results of the model.

[0090] Therefore, according to the training needs, the training is divided into two steps.

[0091] This embodiment sets three transducer array selection maps (amplitude distribution) and one fusion phase plane map (phase control). Figure 3 As shown in Figure 2, after obtaining the output of the network, the three groups of amplitude distributions are combined with the same fusion phase control to calculate the reconstructed sound field where the sound wave is transmitted to the target plane. Its amplitude distribution is: The designed target sound field amplitudes are A1, A2, A3, and the network loss function is constructed based on them.

[0092] In addition, during the training process, if the 8×8 amplitude image is binarized at the beginning of the training, the learning rate needs to be adjusted to a very small value, otherwise nan will appear, that is, the training result will quickly reach the numerical limit. However, if the learning rate is too small at the beginning of the training, the training will be slow and the training effect will be poor. Therefore, the network training is divided into two steps. In the first step, no binarization operation is performed, only normalization operation is performed, and the initial learning rate is set to 10 -3 , train for 10000 epochs;

[0093] In the second step, we continue training based on the training results of the first step, perform binarization, and set the initial learning rate of this step to 10. -6 After completing the training, the results are as follows Figure 3 (b) as shown.

[0094] Finally, the transducer array is controlled by a switch circuit according to the transducer array selection diagram, and the phase holographic plate is arranged according to the phase plane diagram, so that the transducer array can present the target sound field through the phase plate after emitting ultrasonic waves.

[0095] Comparative Example 1

[0096] Iterative angular spectrum method is a common method for designing acoustic holographic sound fields, but it usually only designs the phase, while the amplitude distribution is randomly generated without optimization. The following briefly introduces the steps of iterative angular spectrum method for the design scenario in Example 1:

[0097] Step 1, back propagating from the target sound field plane z=D2+D1 to the phase holographic plate plane z=D1, where the amplitude of the target sound field is A and the phase is 0, and its expression is:

[0098] p3(x,y,z=D1+D2)=A(x,y,z=D1+D2)

[0099] Using the sound propagation model of Example 1, the sound field of the z=D1 phase holographic plate plane is obtained:

[0100]

[0101] Take the phase of p2 as the phase distribution of the phase hologram

[0102] Step 2, randomly generate the amplitude distribution of the transducer array as A1, such as Figure 5 As shown, the sound field is expressed as: p0(x, y, z=0)=A1. Applying the sound propagation model of Example 1, the sound field propagated from the transducer to the plane of the phase holographic plate z=D1 is obtained:

[0103]

[0104] And p′2(x,y,z=D1) and the phase distribution of the phase hologram Combined, the sound field on the z=D1 phase holographic plate plane is obtained as:

[0105] By applying the sound propagation model of Example 1, the sound field distribution of the sound propagating to the target plane z=D2+D1 can be calculated as follows:

[0106]

[0107] Step 3, directly replace the amplitude distribution of p3′(x, y, z=D1+D2) with the target amplitude A to obtain a new target plane sound field p″3, and repeat steps 1 and 2. After 60 iterations, take the phase distribution in step 1 of the last iteration And the amplitude distribution A1 of step 2 is the final result. Figure 6As shown, the result of the iterative angular spectrum method has a fuzzy contour and more artifacts, while the result of the present invention has a clear contour and fewer artifacts. The above compares the embodiment of the present invention with the comparative example. The result of the present invention has a higher contour integrity and fewer artifacts. Unlike the iterative angular spectrum method, which lacks global optimization, the method of the present invention makes up for this defect and has a better design effect.

[0108] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for designing spatial multiplexing acoustic holograms based on a physical model neural network, characterized in that: The following steps are involved: Step 1: Determine various parameters based on the target spatial reuse design scenario and design requirements; Step 2: Establishing a sound wave propagation model, wherein the sound wave propagation model is used to transfer the sound field image of one plane to another plane through the actual scene in the target space multiplexing design scenario, thereby obtaining the sound field image of the other plane; Step 3: Build a deep learning model. The input of the model is a random noise image, and the output is a transducer array selection map and a phase plane map. Constrain the output image to make it meet the parameters preset in step 1. Step 4: Combine the sound wave propagation model and the deep learning model to build an overall model and train it in combination with the target sound field map. After training, the transducer array selection map and phase plane map are obtained for presenting the target sound field map in the target space multiplexing design scenario.

2. The method for designing a spatial multiplexed acoustic hologram based on a physical model neural network according to claim 1, characterized in that: The transducer array is composed of multiple ultrasonic transducers tightly arranged in the same plane, and each ultrasonic transducer has only two states: on and off. The transducer array selection map is a distribution map of the on or off state of each ultrasonic transducer.

3. The method for designing a spatial multiplexed acoustic hologram based on a physical model neural network according to claim 1, characterized in that: The various parameters of step 1 specifically include the spatial excitation area of ​​the transducer array, the length and width of the phase hologram, the production resolution limit of the phase hologram, the image matrix size of the model output, the sound wave propagation distance, frequency, and propagation medium of the transducer array, etc.

4. The method for designing a spatial multiplexed acoustic hologram based on a physical model neural network according to claim 1, characterized in that: The sound wave propagation model of step 2 is specifically: Step 2.1, set the transducer array, the phase hologram plate and the target plane to be spaced in sequence along the positive direction of the z-axis depth direction. The sound pressure field p1 emitted by the transducer array propagates in the positive direction along the z-axis depth direction to the sound pressure field p2 of the phase hologram plate located on the z=D1 plane is: Among them, x and y represent the position variables in the spatial Cartesian coordinate system, k x , k y represent the variables in the wave number space, represents the wave field extrapolation function of Fourier acoustics, p2(x, y, D1) represents the sound pressure field on the z=D1 plane in the spatial Cartesian coordinate system, and k represents the wave number of the sound wave in the sound propagation medium with a wavelength of λ; Step 2.2, the sound field p2 transmitted from the transducer array to the phase hologram undergoes target phase modulation after passing through the phase hologram, so that the transmitted sound field of the phase hologram introduces the phase distribution The sound pressure field p2′ after passing through the phase holographic plate is obtained according to the following formula: Among them, e represents a natural constant, and i represents an imaginary unit; Step 2.3, the modulated sound pressure field p2′ continues to propagate forward along the z-axis depth direction to the target plane located on the z=D2 plane, and the sound pressure field p3 on the target plane is obtained by processing according to the following formula: in, is the wave field extrapolation function of Fourier acoustics.

5. The method for designing spatial multiplexing acoustic holograms based on a physical model neural network according to claim 1, characterized in that: The deep learning model adopts a U-net++ network structure, and performs specific constraint processing on the output image of the U-net++ network structure according to the design of the acoustic holographic device.

6. The method for designing spatial multiplexing acoustic holograms based on a physical model neural network according to claim 1, characterized in that: In step 3, the deep learning model outputs multiple transducer array selection maps / phase plane maps, which are obtained by adjusting the number of output channels in the U-net++ network structure.

7. The method for designing spatial multiplexing acoustic holograms based on a physical model neural network according to claim 1, characterized in that: In step 3, the specific steps are: Step 3.1, using a randomly generated Gaussian noise image of size 512×512 as a random noise image, first use the U-net++ network structure to process the random noise image, and the U-net++ network structure outputs the encoding amplitude distribution A1 corresponding to the transducer array as the transducer array selection map and the phase modulation distribution corresponding to the phase hologram As a phase plane diagram; Step 3.2, then perform an average pooling operation on the transducer array selection image A1, convert the image into 8×8 size, then perform nearest neighbor upsampling to expand the image to 512×512, and finally perform binarization processing; at the same time, perform low-pass filtering on the phase distribution image to filter out the high-frequency components.

8. The method for designing spatial multiplexing acoustic holograms based on a physical model neural network according to claim 1, characterized in that: The step 4 is specifically: 4.

1. First, the transducer array selection map and phase plane map output by the deep learning model are used as the sound field map of the plane in the actual scene, and are input into the sound wave propagation model to obtain the holographic sound field map after the actual scene through calculation; 4.

2. Subtract the holographic sound field image from the target sound field image to obtain the loss; 4.

3. The loss is then input into the deep learning model for back propagation optimization, and the optimized deep learning model outputs the transducer array selection map and phase plane map again; 4.

4. Repeat steps 4.1 to 4.3 for multiple times until the loss reaches the preset loss threshold, and take the transducer array selection map and phase plane map obtained by the final optimization as the output result.

9. The method for designing spatial multiplexing acoustic holograms based on a physical model neural network according to claim 8, characterized in that: In step 4, the loss uses mean square error as the loss function of the network and measures the gap between the hologram optimized by the network and the target hologram. The formula is as follows: in, is the amplitude part of the holographic sound field image derived from the sound wave propagation model, A is the amplitude part of the target holographic sound field, abs() represents the absolute value function, Represents the holographic sound field envelope distribution derived from the sound wave propagation model.

Images