Deep learning-based terahertz superconducting dynamic inductance detector array design method, medium and equipment

Abstract

The invention provides a terahertz superconducting dynamic inductance detector array design method based on deep learning, a medium and equipment, and belongs to the field of terahertz superconducting dynamic inductance detector design. According to the method, a frequency response parameter of the terahertz superconducting KID array is used as an input variable of a deep neural network, a design structure parameter of the terahertz superconducting KID array is used as an output variable of the deep neural network, a mapping data set of the frequency response parameter and the design structure parameter is established, and the deep neural network is trained by using the mapping data set; and inputting a frequency response parameter of the target terahertz superconducting KID array by using the trained deep neural network to obtain a corresponding optimal design structure parameter. According to the method, the deep learning algorithm is combined, the design process of the terahertz superconducting KID array is optimized, the performance and design efficiency of the detector are improved, and the problems of complexity and time consumption of a traditional design method are solved.

Description

Technical Field

[0001] The present invention belongs to the field of terahertz superconducting kinetic inductance detector design, and particularly relates to a design method, medium and device of a terahertz superconducting kinetic inductance detector array based on deep learning. Background Art

[0002] There are mainly two types of superconducting detectors for large-aperture terahertz telescopes: terahertz coherent (heterodyne mixing) superconducting detectors and terahertz incoherent (direct detection) superconducting detectors. Compared with terahertz coherent superconducting detection, terahertz incoherent detection does not require a local oscillator signal, so it has better integration and higher stability, and has developed rapidly in recent years. Representative detectors include Transition Edge Sensor (TES) and Kinetic Inductance Detector (KID). Terahertz superconducting TES detectors and superconducting KID detectors have ultra-high detection sensitivity and can achieve signal detection at the background noise limit. However, the superconducting TES detector uses a time-division multiplexing multi-channel readout circuit based on a superconducting quantum interference current amplifier, that is, each pixel of the TES detector array needs to be equipped with a current amplifier, which is difficult to integrate at low temperature and is not conducive to the development of large-scale array detection technology. Different from the superconducting TES detector, after the superconducting KID detector is irradiated by an external signal, its microwave resonator characteristics (resonance frequency, quality factor Q value, etc.) will change. Therefore, the detection signal information can be obtained by reading the amplitude or phase of the excitation signal at the characteristic frequency of the microwave resonator. In addition, since the microwave resonator can achieve a high-Q value design, the output of a large-scale detector array can be obtained simultaneously only through a transmission line and combined with frequency-division multiplexing readout technology. These characteristics are very conducive to the large-scale system integration of the superconducting KID array.

[0003] However, as the number of pixels increases, the scale of the terahertz superconducting KID array becomes larger and larger. After the structural design and optimization of a single superconducting KID unit are compounded by the array, the complexity of the overall structural design and optimization of the array will increase exponentially. Using electromagnetic simulation algorithms or iterative matching design algorithms mainly based on the finite element difference method will face challenges such as a huge increase in data volume and a huge increase in operation time. In particular, the current design method of the terahertz superconducting KID array is to manually adjust the array structure parameters to obtain different resonance response curves, and the design speed and optimization accuracy are not high enough. Therefore, in the face of the growing demand for large sky area astronomical observations, the high-sensitivity terahertz superconducting KID array faces the problem that the difficulty of array design and optimization increases rapidly as the number of pixels increases.

[0004] Traditional superconducting kinetic inductance detector (KID) array design methods often involve forward design. Although obtaining the frequency response S21 sequence curve from a superconducting KID array is a forward process, i.e., the obtained result is definite and unique, the design process of a superconducting KID array is to first have the desired frequency response parameters and then inversely design the superconducting KID array structure that can generate such frequency response parameters. Therefore, the forward design process highly depends on manual empirical adjustments, and it is very difficult to improve the design speed and optimization accuracy. In the design of terahertz superconducting KID arrays, the process of directly obtaining the array structure parameters from the target resonance response curve is a typical inverse problem solution based on the forward process of Maxwell's equations, and its solution is uncertain and non-unique. Summary of the Invention

[0005] Aiming at the deficiencies in the prior art, the present invention provides a terahertz superconducting kinetic inductance detector array design method, medium, and device based on deep learning. By taking advantage of the characteristics of deep neural networks that do not rely on prior models and manual parameter settings, the present invention is completely driven by the mapping data between the array structure parameters and the target resonance response curve. Taking the desired frequency response parameters as the input and the structure parameters as the output, and using the error between the optimal design structure parameters and the design structure parameters output by the deep neural network as the objective function for constraint, the corresponding deep neural network model is trained to complete the rapid and automatic design of the large-scale array structure of terahertz superconducting KIDs.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In the first aspect, the present invention provides a terahertz superconducting kinetic inductance detector array design method based on deep learning, including: taking the frequency response parameters of a terahertz superconducting KID array as the input variables of a deep neural network, taking the design structure parameters of a terahertz superconducting KID array as the output variables of the deep neural network, establishing a mapping data set between the frequency response parameters and the design structure parameters, and training the deep neural network using the mapping data set; using the trained deep neural network, inputting the frequency response parameters of the target terahertz superconducting KID array, and obtaining the corresponding optimal design structure parameters.

[0008] Optionally, the deep neural network adopts a hybrid deep neural network including a fully connected layer, a convolutional layer, and a normalization layer.

[0009] Optionally, the terahertz superconducting KID array is a microstrip KID array, including an Nb thin film, a high-resistivity silicon substrate, and an Nb thin film substrate stacked in sequence from top to bottom.

[0010] Optionally, the frequency response parameter is the S21 sequence curve of the terahertz superconducting KID array, and the design structure parameters are the Nb film height H1, the high-resistivity silicon substrate height H2, the Nb film substrate height H3, the Nb film width W1, the Nb film substrate width W2, and the Nb film microstrip line length L.

[0011] Optionally, the mapping data set is established by a forward process based on Maxwell's equations. The specific process is as follows:

[0012] According to the design structure parameters H1, H2, H3, W1, W1, and L, seven characteristic parameters of the KID array are obtained through a vector electromagnetic simulation algorithm, namely the total gain a of the KID array unit, the system phase shift α, the system link delay τ, the load quality factor Qr of the KID array unit, the coupling quality factor Qc, the resonance frequency fr, and the phase angle generated by impedance mismatch.

[0013] The S21 sequence curve is obtained according to the forward coupling formula of the seven characteristic parameters. The formula is as follows:

[0014]

[0015] In the formula, S 21 (f) represents the S21 sequence curve data, and f represents the frequency.

[0016] Optionally, in the process of training the deep neural network using the mapping data set, the objective function used is as follows:

[0017]

[0018] In the formula, loss represents the objective function, X i represents the set composed of six design structure parameters in the mapping data set, and X′ i represents the set composed of six design structure parameters currently output by the deep neural network.

[0019] Optionally, the specific process of training the deep neural network using the mapping data set is as follows:

[0020] Rearrange the mapping data set. Use the S21 sequence curve data in the mapping data set as the input of the deep neural network, and use the design structure parameters in the mapping data set as the output of the deep neural network. Train the deep neural network until the network converges.

[0021] Optionally, in the process of training the deep neural network using the mapping data set, the real part and the imaginary part of the S21 sequence curve data are extracted and combined into a fully real sequence for input.

[0022] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program, which causes a computer to execute the method for designing a terahertz superconducting kinetic inductance detector array based on deep learning as described in the first aspect.

[0023] In a third aspect, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method for designing a terahertz superconducting kinetic inductance detector array based on deep learning as described in the first aspect is implemented.

[0024] The beneficial effects of the present invention are as follows: In view of the fact that traditional inverse design mainly relies on an iterative matching method with additional prior information to obtain an approximate solution, which often faces problems such as complex iterative calculations, long time consumption, and low accuracy, the present invention proposes a method for rapidly designing a large-scale array of terahertz superconducting kinetic inductance detectors based on deep learning. The present invention makes full use of the advantages of deep learning, and uses the error between the optimal design structure parameters and the design structure parameters output by the deep neural network as the objective function constraint to train the corresponding deep neural network. During the design process, no manual adjustment and additional parameter settings are required. By inputting the frequency response parameters of the desired superconducting KID array, the trained deep neural network can directly and rapidly output the optimal design structure parameters of the superconducting KID array, enabling efficient automatic inverse design of the superconducting KID array and greatly reducing the difficulty and time consumption of designing the superconducting KID array. Description of the Drawings

[0025] Figure 1 is a schematic diagram of the method for designing a terahertz superconducting kinetic inductance detector array based on deep learning.

[0026] Figure 2 is an architecture diagram of a deep neural network for inverse design of the structural parameters of a KID array unit.

[0027] Figure 3 is a curve graph showing the change of the composite loss value with the number of training steps during the training process. Detailed Embodiments

[0028] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application.

[0029] In one embodiment, the present invention proposes a method for designing a terahertz superconducting kinetic inductance detector array based on deep learning. As Figure 1As shown in the figure, the basic principle of the design method of a terahertz superconducting dynamic inductance detector array based on deep learning is as follows: taking the frequency response parameters of a terahertz superconducting KID array as the input variables of a deep neural network, taking the design structure parameters of a terahertz superconducting KID array as the output variables of the deep neural network, establishing a mapping data set of frequency response parameters and design structure parameters, and using the mapping data set to train the deep neural network; using the trained deep neural network, inputting the frequency response parameters of the target superconducting KID array to obtain the corresponding optimal design structure parameters.

[0030] In this embodiment, a 1.4T terahertz microstrip-type silicon-based Nb superconducting thin film KID array is taken as the design target, and the implementation steps are as follows:

[0031] (1) According to the data structure characteristics of the input and output, the deep learning network architecture adopts a hybrid deep neural network including a fully connected layer, a convolutional layer, and a normalization layer. Such a network architecture has good multi-dimensional data feature extraction capabilities and can efficiently approximate the mapping function of structural parameter design. The specific network architecture is as Figure 2 shown.

[0032] (2) The specific input and output parameters of the deep neural network are as Figure 1 shown. The input is the S21 sequence curve of the target KID array unit, and the output is the Nb thin film height H1, the high-resistance silicon substrate height H2, the Nb thin film substrate height H3, the Nb thin film width W1, the Nb thin film substrate width W2, and the Nb thin film microstrip line length L. In this embodiment, a microstrip-type KID array is adopted.

[0033] (3) The mapping data set is established by a forward process based on Maxwell's equations. First, seven main characteristic parameters that can describe the characteristics of the KID array unit are obtained from the structural parameters H1, H2, H3, W1, W1, and L through a vector electromagnetic simulation algorithm. They are the total gain a of the KID array unit, the system phase shift α, the system link delay τ, the load quality factor Qr of the KID array unit, the coupling quality factor Qc, the resonance frequency fr, and the phase angle generated by impedance mismatch Then, the S21 sequence curve data is obtained from the forward coupling formula of these seven characteristic parameters. The specific formula is as follows:

[0034]

[0035] In the formula, f represents the frequency.

[0036] The establishment of the entire mapping data set from the design structure parameters to the main characteristic parameters and then to the S21 sequence curve data is a forward process. Therefore, the establishment process of the mapping data set is determined and unique.

[0037] (4) Establish a composite objective function. For a deep neural network, there are significant differences in the influence degree of each parameter in the structural parameters of the KID array unit on the S21 sequence curve data. Therefore, a composite loss composed of the L1-norm error and the L2-norm error is used as the objective function, and the specific formula is as follows:

[0038]

[0039] In the formula, X i represents the set composed of six design structure parameters, and X′ i represents the set composed of six design structure parameters currently output by the deep neural network. The first term is the L1-norm error term, which is strengthened by 10 times to balance the different influences of each parameter on the S21 sequence curve data. The second term is the L2-norm error term.

[0040] (5) According to step (3), establish a mapping dataset of 60,000 groups from design structure parameters to S21 sequence curve data, and rearrange the mapping dataset. Take the S21 sequence curve data in the mapping dataset as the input of the deep neural network, and take the structure parameters in the mapping dataset as the output of the deep neural network. Here, since the S21 sequence curve data is a complex value (n×1), and the deep neural network only supports real number training, the real part and the imaginary part of the S21 sequence curve data are respectively extracted to form a full real number sequence (2n×1). Train the designed deep neural network with the rearranged mapping dataset until the network converges. The change curve of the composite loss value with the number of training steps is as Figure 3 shown. As Figure 3 can be seen, using the pre-established 60,000 groups of training datasets, the deep neural network quickly completed convergence and met the requirements of inverse design.

[0041] (5) Obtain the deep neural network for the single microstrip silicon-based Nb superconducting thin film KID array unit structure for inverse design, and fix all the parameters in the network as the terahertz superconducting KID array design network.

[0042] (6) Input any desired S21 sequence curve into the terahertz superconducting KID array design network to obtain the optimal design structure parameters of the KID array unit.

[0043] (7) Summarize the optimal design structure parameters of all KID array units to obtain all the optimal design structure parameters of the entire array.

[0044] In this embodiment, a data-driven deep neural network is used to efficiently and rapidly approximate and solve the mapping function of the input and output data spaces, successfully expanding the application scope of deep learning from feature extraction and recognition to a series of inverse problem solutions, especially the inverse problem solution based on the forward process of Maxwell's equations, meeting the rapid design requirements of large-scale terahertz superconducting KID arrays.

[0045] In another embodiment, the present invention proposes a computer-readable storage medium storing a computer program, which causes a computer to execute the deep learning-based terahertz superconducting kinetic inductance detector array design method of 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, wherein when the processor executes the computer program, the deep learning-based terahertz superconducting kinetic inductance detector array design method of the foregoing embodiment is implemented.

[0047] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that can 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 disk 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 is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. Any technical solution falling within the concept of the present invention belongs to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, 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 design method for a terahertz superconducting dynamic inductance detector array based on deep learning, characterized in that, Including: The frequency response parameters of the terahertz superconducting KID array are used as the input variables of the deep neural network, and the design structure parameters of the terahertz superconducting KID array are used as the output variables of the deep neural network. A mapping data set between the frequency response parameters and the design structure parameters is established, and the deep neural network is trained using the mapping data set. Using the trained deep neural network, the frequency response parameters of the target terahertz superconducting KID array are input to obtain the corresponding optimal design structure parameters.

2. The design method of a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 1, characterized in that: The deep neural network adopts a hybrid deep neural network including a fully connected layer, a convolutional layer, and a normalization layer.

3. The design method of a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 1, characterized in that: The terahertz superconducting KID array is a microstrip KID array, including an Nb thin film, a high-resistivity silicon substrate, and an Nb thin film substrate stacked in sequence from top to bottom.

4. The design method of a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 3, characterized in that: The frequency response parameters are the S21 sequence curve of the terahertz superconducting KID array, and the design structure parameters are the height H1 of the Nb thin film, the height H2 of the high-resistivity silicon substrate, the height H3 of the Nb thin film substrate, the width W1 of the Nb thin film, the width W2 of the Nb thin film substrate, and the length L of the Nb thin film microstrip line.

5. The design method of a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 4, wherein: The mapping data set is established by a forward process based on Maxwell's equations. The specific process is as follows: According to the design structure parameters H1, H2, H3, W1, W1, and L, seven characteristic parameters of the KID array are obtained through the vector electromagnetic simulation algorithm, which are the total gain a of the KID array unit, the system phase shift α, the system link delay τ, the load quality factor Qr of the KID array unit, the coupling quality factor Qc, the resonant frequency fr, and the phase angle generated by impedance mismatch The S21 sequence curve is obtained according to the forward coupling formula of seven characteristic parameters. The formula is as follows: Where S 21 (f) represents the S21 sequence curve data, and f represents the frequency.

6. The design method of a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 4, characterized in that: During the process of training the deep neural network using the mapping data set, the objective function adopted is as follows: where loss represents the objective function, X i represents the set composed of six design structure parameters in the mapping dataset, and X′ i represents the set composed of six design structure parameters currently output by the deep neural network.

7. The method for designing a terahertz superconducting dynamic inductance detector array based on deep learning according to claim 4, wherein: The specific process of training the deep neural network using the mapping data set is as follows: The mapping data set is rearranged. The S21 sequence curve data in the mapping data set is used as the input of the deep neural network, and the design structure parameters in the mapping data set are used as the output of the deep neural network. The deep neural network is trained until the network converges.

8. The design method of a terahertz superconducting kinetic inductance detector array based on deep learning according to claim 7, characterized in that: During the process of training the deep neural network using the mapping data set, the real part and the imaginary part of the S21 sequence curve data are extracted to form a fully real sequence for input.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the method for designing a terahertz superconducting dynamic inductance detector array based on deep learning according to any one of claims 1-8.

10. 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, the method for designing a terahertz superconducting dynamic inductance detector array based on deep learning according to any one of claims 1-8 is implemented.

Images