Terahertz grating design method based on deep neural network and optical diffraction model, medium and equipment
By combining deep neural networks and optical diffraction models, the target output and actual output of the terahertz grating are directly compared, and the problem of local optimal solution in traditional methods is solved, and an efficient global optimal terahertz grating design is achieved.
Patent Information
- Application Number
- CN202510418403.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-03
- Publication Date
- 2025-07-18
AI Technical Summary
Traditional terahertz grating design methods are prone to falling into local optimal solutions in complex and large-scale designs, and the data acquisition cost is high, making it difficult to find the global optimal solution.
Combining the deep neural network and optical diffraction model, the target output parameters of the terahertz grating are directly compared with the output results of the optical vector diffraction model, the objective function is constructed, and the global optimal solution is found through the self-cycle training process.
There is no need to pre-establish data mapping sets, improve design efficiency, avoid local optimal solutions, and achieve more efficient terahertz grating design.
Smart Images
Figure CN120335154A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of terahertz grating design, and particularly relates to a terahertz grating design method, medium and device based on a deep neural network and an optical diffraction model. Background Art
[0002] Terahertz waves are located between microwaves and infrared rays in the electromagnetic spectrum, with a frequency range usually between 0.1 and 10 THz. They have unique physical properties and extensive application potential, such as in the fields of imaging, communication, material detection, and biomedicine. A terahertz grating is an optical element used to precisely modulate and control the propagation and properties of terahertz waves. Its design and manufacturing technologies involve various materials and structures, including metal gratings, dielectric gratings, and nano gratings, etc. Among them, metal gratings usually have high reflectivity and good diffraction characteristics, while dielectric gratings can adjust the propagation characteristics of terahertz waves by changing the refractive index of materials. With the rapid development of terahertz technology, gratings, as an important optical element, are often used to construct terahertz spectrometers, terahertz imaging systems, and other important terahertz spectroscopic imaging devices. In the field of astronomy, terahertz grating technology is also playing an increasingly important role in the field of terahertz astronomical observation technology. It can provide multi-beam and large-field-of-view observation effects, and can greatly improve the observation efficiency and observation sensitivity for large-area celestial targets.
[0003] Traditional terahertz grating design methods usually first establish a mathematical model describing the performance of terahertz Fourier gratings, that is, establish a system of linear equations using physical properties such as terahertz propagation, diffraction, and interference, and then seek the solution of the system of linear equations through an iterative method. Finally, the solution of the system of linear equations is gradually approximated through this iterative method. In this process, the result of the current iteration is usually used to update the value of the next variable until a stable solution is converged. This iterative method can be used to optimize the geometric parameters (such as period, depth, and material properties) of the grating in the design of terahertz gratings to meet the best diffraction efficiency and performance requirements. The advantage of this method is that the iteration speed is relatively fast when dealing with a small-scale system of linear equations. However, for some complex and large-scale grating designs, the iterative method may have convergence problems, especially when the initial value is not selected properly, and it is very easy to fall into a local optimal solution and fail to find the global optimal solution.
[0004] With the introduction of advanced technologies such as deep learning, the design and optimization process of terahertz gratings is also constantly evolving. Through data-driven methods, deep learning can more efficiently find the optimal solution, thus achieving higher-performance terahertz grating designs. However, in deep learning methods, the training of deep neural networks often relies on the establishment of large-scale datasets. In the fields of image recognition and natural language processing, the difficulty and cost of obtaining and establishing large-scale datasets of data such as images and languages are relatively low. Therefore, relying on these large-scale datasets can train deep neural network models with excellent performance for image recognition and language processing. However, for terahertz grating design, the mapping model between grating structure parameters and grating output parameters has unacceptable data acquisition difficulty and data acquisition cost, and for each specific terahertz frequency, data needs to be re-acquired and a mapping model needs to be established, which is difficult to be practical. Summary of the Invention
[0005] In view of the deficiencies in the prior art, the present invention provides a terahertz grating design method, medium, and device based on a deep neural network and an optical diffraction model. The present invention does not require the pre-establishment of any terahertz grating parameter mapping model, directly combines the training of the deep neural network and the optical vector diffraction model; at the same time, it does not use the difference between the model output and the optimization target in traditional deep learning methods as the objective function, but directly compares the output result of the model input into the optical vector diffraction model with the pre-given target design result as the objective function, overcoming the problem that traditional deep learning itself is prone to falling into local optima and making it easier to find the global optimal solution.
[0006] To achieve the above object, the present invention adopts the following technical solutions:
[0007] In a first aspect, the present invention provides a terahertz grating design method based on a deep neural network and an optical diffraction model, including: using the target output parameters of the terahertz grating as the input variables of the deep neural network, using the structure parameters of the terahertz grating as the output variables of the deep neural network, and constructing the deep neural network; inputting the output result of the deep neural network into the optical vector diffraction model, comparing the output parameters of the optical vector diffraction model with the target output parameters of the terahertz grating to establish an objective function, and starting the training of the deep neural network based on the objective function; using the trained deep neural network, inputting the target output parameters of the terahertz grating, and obtaining the optimal structure parameters of the corresponding terahertz grating.
[0008] Optionally, the deep neural network adopts a hybrid deep neural network containing only fully connected layers and convolutional layers.
[0009] Optionally, the target output parameter of the terahertz grating is the power intensity distribution of the terahertz beam formed after reflection and diffraction by the terahertz n×n metal grating array, which is represented by an m×m matrix containing only real numbers, and each element in the matrix represents the power intensity of the terahertz beam at that point.
[0010] Optionally, the structural parameters of the terahertz grating are represented by an n×n matrix containing only real numbers, and each element in the matrix represents the height of the terahertz metal grating at that point.
[0011] Optionally, the optical vector diffraction model calculates the current actual power intensity distribution of the terahertz beam after grating diffraction according to the output result of the deep neural network.
[0012] Optionally, the formula of the objective function is:
[0013] Loss=L2(X, X′)+0.05(1 - SSIM(X, X′))
[0014] Where
[0015]
[0016] In the formula, Loss represents the objective function, X represents the target terahertz beam power intensity distribution, X′ represents the current actual terahertz beam power intensity distribution, X i represents the target terahertz beam power intensity at the i-th point, X′ i represents the current actual terahertz beam power intensity at the i-th point, u X and u X′ are the average values of X and X′ respectively, σ X and σ X′ are the standard deviations of X and X′ respectively, σ XX′ is the covariance of X and X′, and C1 and C2 are constants.
[0017] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program, and the computer program causes a computer to execute the terahertz grating design method based on a deep neural network and an optical diffraction model as described in the first aspect.
[0018] In a third aspect, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, it implements the terahertz grating design method based on a deep neural network and an optical diffraction model as described in the first aspect.
[0019] The beneficial effects of the present invention are:
[0020] 1. Low cost and strong adaptability: After calculating the output result of the deep neural network through the optical vector diffraction model, the present invention constructs an objective function together with the desired grating output result and re-enters the training process of the deep neural network, constructing a self-loop process. Therefore, the present invention does not need to pre-collect and establish any data mapping sets, and directly starts the entire calculation process from the desired grating output result, so it can adapt to the application requirements of different terahertz frequencies;
[0021] 2. Improve design efficiency: By combining deep learning and the optical vector diffraction model, in each round of calculation cycle of the present invention, the output result of each round of the deep neural network is directly compared with the desired grating output result finally, and the comparison result is used as the driving force for updating the parameters of the deep neural network. The result orientation of the process is obvious, saving a large amount of additional auxiliary parameter calculations and storage spaces, and can significantly improve the design efficiency of terahertz gratings, reducing the cumbersome iterative calculation process in traditional design methods;
[0022] 3. Overcome the local optimum problem: The self-loop training design process of the present invention makes the optimization process of the deep neural network always take the desired grating output result set in advance as the final and only orientation. At the same time, it combines the advantages of large-scale parameter generalization and search of the deep neural network, searching in a large range in all possible solution set spaces, achieving the combination of the final only orientation goal and large-scale parameter search. Therefore, it is easier to find the global optimum solution, avoiding the risk of falling into the local optimum solution in traditional iterative methods and deep learning methods, thereby improving the accuracy and reliability of the design. Description of the Drawings
[0023] Figure 1 is a flowchart of a terahertz grating design method based on a deep neural network and an optical diffraction model.
[0024] Figure 2a and 2b are respectively two-dimensional and three-dimensional distribution schematic diagrams of the terahertz beam power intensity distribution data output by a terahertz metal grating array for realizing the beam splitting function.
[0025] Figure 3 is a three-dimensional data distribution schematic diagram of the actual size parameters of a terahertz metal grating array for realizing the beam splitting function.
[0026] Figure 4 is a schematic diagram of the forward process of obtaining the actual terahertz beam splitting intensity distribution after grating diffraction from the actual size parameters of a terahertz metal grating using the optical vector diffraction model.
[0027] Figure 5 is a schematic diagram of the reverse process of obtaining the actual size parameters of a terahertz metal grating from the desired terahertz beam splitting intensity distribution.
[0028] Figure 6 It is a curve graph showing the change of the specific value of the objective function Loss with the number of cycle steps. DETAILED DESCRIPTION
[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.
[0030] In one embodiment, the present invention proposes a terahertz grating design method based on a deep neural network and an optical diffraction model, the basic principle of which is: taking a predetermined desired terahertz grating output parameter as an input variable, and taking a structural parameter of the terahertz grating as an output variable, inputting the output result of the deep neural network into the terahertz optical vector diffraction model and directly comparing it with the predetermined desired design result as an objective function, starting the training of the deep neural network, and after the objective function reaches a specified minimum accuracy, the obtained network output is the optimal structural parameter of the terahertz grating.
[0031] This embodiment takes a terahertz n×n metal grating array for dividing a terahertz wave beam into four beams as a design target, and the overall implementation steps of the method are as follows:
[0032] (1) The output parameter of the terahertz metal grating is the far-field terahertz beam power intensity distribution formed after the terahertz beam is reflected and diffracted by the terahertz metal grating. Therefore, it is represented by an m×m matrix containing only real numbers. Each element in the matrix represents the actual terahertz wave power beam power intensity at that point, such as Figure 2a and 2b As shown in the figure, it is a two-dimensional data representation and a three-dimensional data representation of a preset terahertz metal grating output parameter distribution. The terahertz metal grating output parameter is used as the input of the deep neural network. The structural parameters of the terahertz n×n metal grating array are represented by an n×n matrix containing only real numbers. Each parameter in the matrix represents the actual terahertz metal grating height at that point, such as Figure 3 As shown in FIG. 1 , a three-dimensional data representation of the parameters of a terahertz metal grating array is shown. The goal of this embodiment is to Figure 2a and 2b Preset the desired terahertz wave power beam power intensity distribution and obtain Figure 3 The specific structural dimensions of the terahertz metal grating array shown in the figure are used to process such a terahertz metal grating array. The processed terahertz metal grating array is used to realize the actual function of dividing a beam of terahertz waves into four beams. According to the data structure characteristics of input and output, the deep neural network architecture adopts a hybrid deep neural network containing only fully connected layers and convolutional layers. This type of network architecture has good data transmission characteristics and can realize the efficient establishment of high-dimensional nonlinear data models.
[0033] (2) The specific input and output parameters of the deep neural network are as follows Figure 1 shown. The input is the target terahertz grating parameter 1, denoted by X, and the output is the n×n terahertz grating structure parameter 3, denoted by Y.
[0034] (3) After the grating structure parameter 3 is input into the optical vector diffraction model 4, through the calculation of the optical vector diffraction model 4, the current actual terahertz beam power intensity distribution parameter after grating diffraction is obtained, denoted by X'. This process is a unique and definite forward process that can obtain a unique result, as Figure 4 shown. The process of obtaining the terahertz grating structure parameter Y from the target terahertz beam power intensity distribution parameter X is a non-unique and indefinite reverse process, as Figure 5 shown, which is the goal to be achieved in this embodiment.
[0035] (4) The objective function 5 is established jointly using the current actual terahertz beam power intensity distribution parameter X' and the target terahertz grating parameter X. The objective function Loss is composed of the L2 norm error and the structural similarity error between the current actual terahertz beam power intensity distribution parameter X' and the target terahertz grating parameter X. The specific formula is as follows:
[0036] Loss = L2(X, X') + 0.05(1 - SSIM(X, X'))
[0037] where
[0038]
[0039] in the formula, u X and u X′ are the average values of X and X' respectively, σ X and σ X′ are the standard deviations of X and X' respectively, σ XX′ is the covariance of X and X', and C1 and C2 are constants.
[0040] (5) The parameters of the designed deep neural network 2 are trained using the objective function 5.
[0041] (6) After all settings are completed, the training is started by the target terahertz grating parameter 1, passing through the deep neural network 2, the terahertz grating structure parameter 3, the optical vector diffraction model 4, the objective function 5 in sequence, and then entering the deep neural network 2 again for closed-loop training.
[0042] (7) When the objective function 5 reaches the pre-set accuracy, the training is stopped.
[0043] (8) The terahertz grating structure parameter 3 obtained by using the trained deep neural network 2 is the grating structure parameter with the optimal data.
[0044] As Figure 6 shown, from the curve of the specific value of the objective function Loss changing with the number of loop steps in this embodiment, it can be seen that as the number of loop steps increases, the Loss value shows an overall downward trend, and the true global optimal value is gradually found. The change curve can also clearly reflect that during the loop process, when Loss falls into the current local optimal value, after the deep neural network jumps out of the current local optimal through a large range of parameter generalization and searches in the wrong direction, due to the Loss constraint with a clear guiding direction, Loss will drop again and search in the correct direction, realizing gradually jumping out of the current local optimal value in this embodiment and efficiently approaching the true global optimal value.
[0045] In another embodiment, the present invention proposes a computer-readable storage medium storing a computer program, and the computer program enables a computer to execute the terahertz grating design method based on a deep neural network and an optical diffraction model in the foregoing embodiment.
[0046] In another embodiment, the present invention proposes an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the terahertz grating design method based on a deep neural network and an optical diffraction model in the foregoing embodiment is implemented.
[0047] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that may contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing. More specific examples of the computer storage medium would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disc read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0048] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.
[0049] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. Any technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.
Claims
1. A terahertz grating design method based on a deep neural network and an optical diffraction model, characterized in that Including: Using the target output parameters of the terahertz grating as the input variables of the deep neural network and the structural parameters of the terahertz grating as the output variables of the deep neural network to construct a deep neural network; inputting the output result of the deep neural network into the optical vector diffraction model, comparing the output parameters of the optical vector diffraction model with the target output parameters of the terahertz grating to establish an objective function, and starting the training of the deep neural network based on the objective function; using the trained deep neural network, inputting the target output parameters of the terahertz grating, and obtaining the optimal structural parameters of the corresponding terahertz grating.
2. The terahertz grating design method based on a deep neural network and an optical diffraction model according to claim 1, wherein: The deep neural network adopts a hybrid deep neural network containing only fully connected layers and convolutional layers.
3. The terahertz grating design method based on a deep neural network and an optical diffraction model according to claim 1, wherein: The target output parameters of the terahertz grating are the terahertz beam power intensity distribution formed after the terahertz beam is reflected and diffracted by the terahertz n×n metal grating array, which is represented by an m×m matrix containing only real numbers, and each element in the matrix represents the terahertz beam power intensity at that point.
4. The terahertz grating design method based on a deep neural network and an optical diffraction model according to claim 3, wherein: The structural parameters of the terahertz grating are represented by an n×n matrix containing only real numbers, and each element in the matrix represents the height of the terahertz metal grating at that point.
5. The terahertz grating design method based on a deep neural network and an optical diffraction model according to claim 3, characterized in that: The optical vector diffraction model calculates the current actual terahertz beam power intensity distribution after grating diffraction according to the output result of the deep neural network.
6. The terahertz grating design method based on a deep neural network and an optical diffraction model according to claim 3, characterized in that: The formula of the objective function is: Loss=L2(X,X′)+0.05(1-SSIM(X,X′)) Where In the formula, Loss represents the objective function, X represents the target terahertz beam power intensity distribution, X′ represents the current actual terahertz beam power intensity distribution, X i represents the target terahertz beam power intensity at the i-th point, X′ i represents the current actual terahertz beam power intensity at the i-th point, u X and u X′ are the average values of X and X′ respectively, σ X and σ X′ are the standard deviations of X and X′ respectively, σ XX′ is the covariance of X and X′, and C1 and C2 are constants.
7. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the terahertz grating design method based on the deep neural network and the optical diffraction model according to any one of claims 1-6.
8. An electronic device, characterized in that, Including: A memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the terahertz grating design method based on the deep neural network and the optical diffraction model according to any one of claims 1-6.