Passive structural design for phase antenna arrays

Abstract

A system and technique for generating passive structure designs are disclosed. A passive structure configured according to the generated design improves the performance of a phase-controlled antenna array into which the passive structure is incorporated. One device includes a non-temporary computer-readable storage medium on which instructions are encoded, and a processor coupled to the non-temporary computer-readable storage medium. The processor executes the instructions to receive the passive structure design, uses a simulation technique to generate metrics based on the passive structure design and frequency data, generates a pool of samples, uses an acquisition function to select a final sample from the pool of samples, provides the final sample to the simulation technique to generate metrics, where the final sample is associated with a new passive structure and configured to output the new passive structure.

Description

[Background technology]

[0001] This disclosure relates to a system and technology for designing passive structures used in phase antenna arrays. [Overview of the Initiative]

[0002] As part of upgrading the current mobile communications network infrastructure to provide 5G voice and data services, millimeter-wave (mmWave) phase-controlled array antennas are being installed at existing radio access network (RAN) cell sites. As used herein, “5G” refers to voice and data services compliant with the fifth-generation technology standard for broadband cellular networks. Each such site typically supports a three-sector antenna array, with each antenna array providing 120° (i.e., ±60°) azimuth coverage within a given cell. These three sector antennas combined provide 360° omnidirectional coverage around the site. Typical 4G / LTE systems operate at frequencies below 6 GHz, resulting in lower propagation loss compared to millimeter-wave frequencies (typically above 20 GHz). To provide network coverage over a range (distance from the RAN cell site) comparable to typical 4G / LTE systems, millimeter-wave antennas must (i) have high directivity and (ii) be beam-steering capable of radiating patterns. Antenna arrays containing multiple radiating elements provide improvements in these performance aspects in the context of millimeter-wave equipment. Antenna arrays generally provide spatial diversity (i.e., the ability for a base station to communicate with multiple devices within the same cell using highly directional antennas, utilizing the same time-frequency resources) through the massive multiple-input multiple-output (massive-MIMO) architecture in 5G system specifications.

[0003] However, high directivity leads to one or more performance degradations. Achieving high directivity requires a large number of elements in each phase control array, and the beam spreads as the phase control array transmits off the optical axis (called the "antenna boresight"), limiting the azimuth scan range of the entire antenna system. Therefore, because millimeter-wave phase control arrays have high directivity, they cannot provide 120° coverage and cause significant gain degradation at wide scan angles. This gain degradation leads to reduced coverage at sector seams within the communication cell.

[0004] The technology of this disclosure relates to the design of a passive structure (e.g., a dielectric lens) for extending the scan range of a millimeter-wave phase-controlled antenna array. The potential advantages of the passive structure design (e.g., dielectric lens design) of this disclosure relate to narrowing the beam width. For example, a low-order array (i.e., an array with a small number of antenna elements) incorporating the passive structure design of this disclosure can provide resolution equivalent to that of a high-order array (i.e., an array with a large number of antenna elements).

[0005] As an example, a passive structure generator includes at least one non-temporary computer-readable storage medium and one processor coupled to the at least one non-temporary computer-readable storage medium. Instructions are stored in the at least one non-temporary computer-readable storage medium. The at least one processor is configured to execute instructions that perform the following: receiving a passive structure design; generating a figure of merit using a simulation method based on the passive structure design and frequency data; generating a pool of samples; selecting a final sample from the pool of samples using an acquisition function; providing the final sample to the simulation method to generate a figure of merit, where the final sample is associated with a new passive structure; and outputting the new passive structure to at least one of a user interface, an external device, or the at least one non-temporary computer-readable storage medium.

[0006] As another example, the device includes interface hardware, memory hardware, and processing circuitry communicatively connected to the memory hardware. The interface hardware is configured to receive passive structure designs. The memory hardware is configured to store passive structure designs received via the interface hardware. The processing circuitry is configured to perform the following based on the passive structure designs and frequency data stored in the memory hardware: generate a figure of merit using a simulation method, generate a pool of samples, select a final sample from the pool of samples using an acquisition function, provide the final sample to the simulation method to generate a figure of merit, where the final sample is associated with a new passive structure, and output the new passive structure to a user interface, an external device, or at least one non-temporary computer-readable storage medium. The processing circuitry is also configured to output the new passive structure design via the interface hardware, or / or store the new passive structure design in the memory hardware.

[0007] As another example, the method includes: a processing circuit of a computing device receiving a passive structure design; a processing circuit generating a figure of merit using a simulation method based on the passive structure design and frequency data; a processing circuit generating a pool of samples; a processing circuit selecting a final sample from the pool of samples using an acquisition function; a processing circuit providing the final sample to a simulation method to generate a figure of merit, wherein the final sample is associated with a new passive structure; and a processing circuit outputting the new passive structure to a user interface, an external device, or at least one non-temporary computer-readable storage medium.

[0008] As another example, the apparatus includes: means for receiving a passive structure design; means for generating a figure of merit using a simulation method based on the passive structure design and frequency data; means for generating a pool of samples; means for selecting a final sample from the pool of samples using an acquisition function; means for providing the final sample to a simulation method to generate a figure of merit, wherein the final sample is associated with a new passive structure; and means for outputting the new passive structure to a user interface, an external device, or at least one non-temporary computer-readable storage medium.

[0009] As another example, a non-temporary computer-readable storage medium is encoded in instructions. When executed by a processing circuit of a computer, these instructions cause the processing circuit to perform the following actions: receive a passive structure design; generate a figure of merit using a simulation method based on the passive structure design and frequency data; generate a pool of samples; select a final sample from the pool of samples using an acquisition function; provide the final sample to the simulation method to generate a figure of merit, where the final sample is associated with a new passive structure; and output the new passive structure to a user interface, an external device, or at least one non-temporary computer-readable storage medium.

[0010] The passive structure design techniques of this disclosure provide several technical improvements in the field of phase-controlled antenna array design. In this way, the passive structure designs of this disclosure can reduce the complexity and cost of designing phase-controlled antenna array systems by reducing the number of one or more phase shifters, amplifiers, and / or impedance matching networks required in the system. Passive structures designed according to the techniques of this disclosure provide these performance improvements in the context of 3-sector antenna implementations, but in many cases the infrastructure can be reduced to a 1-antenna or 2-antenna array, especially in use cases covering smaller, densely populated areas such as urban centers. Furthermore, the improvements provided by passive structures designed according to the techniques of this disclosure can improve performance and capacity at the cell level and thus reduce the number of arrays required from a higher-order (e.g., system level) perspective. [Brief explanation of the drawing]

[0011] [Figure 1A] Figures 1A and 1B illustrate the difference between the enhanced signal coverage provided by a corresponding phase antenna array with a passive structure designed according to the technology of this disclosure and the signal coverage provided by a conventional phase antenna array. [Figure 1B] Figures 1A and 1B illustrate the difference between the enhanced signal coverage provided by a corresponding phase antenna array with a passive structure designed according to the technology of this disclosure and the signal coverage provided by a conventional phase antenna array.

[0012] [Figure 2] Figure 2 is a block diagram showing an example of an implementation of the passive structure generating device of this disclosure.

[0013] [Figure 3] Figure 3 is a flowchart of an example workflow in this disclosure.

[0014] [Figure 4] Figure 4 is a flowchart showing an example process that the passive structure generation device of FIG. 2 can implement to execute the Bayesian optimization technique of the present disclosure.

[0015] [Figure 5] Figure 5 is a graph showing the clustering of a sample set within the constraints imposed according to the technique of the present disclosure.

[0016] [Figure 6A] Figure 6A is a graph showing the convergence of the figure of merit (FoM) over multiple executions of an electromagnetic field simulator (EM) for an expected improvement (EI) function located at the search end of an acquisition function spectrum. [Figure 6B] Figure 6B is a graph showing the distribution of width and pitch for the EI acquisition function related to FIG. 6A.

[0017] [Figure 7] Figure 7 is a graph showing the convergence of the figure of merit (FoM) over multiple executions of an electromagnetic field simulator.

[0018] [Figure 8] Figure 8 is a graph showing the convergence of the figure of merit (FoM) over multiple executions of an electromagnetic field simulator.

[0019] [Figure 9A] Figure 9A is a diagram showing various dielectric lens shapes generated by the passive structure generation device of FIG. 2 according to the Bayesian optimization design technique of the present disclosure. [Figure 9B] Figure 9B is a diagram showing various dielectric lens shapes generated by the passive structure generation device of FIG. 2 according to the Bayesian optimization design technique of the present disclosure. [[ID=4)]

[0020] [Figure 10] Figure 10 is a graph showing an example of the performance metrics (i.e., peak gain) of various phase antenna arrays at various scan angles.

[0021] [Figure 11] Figure 11 shows aspects of various optimized passive structural designs of this disclosure compared to conventional passive structural designs.

[0022] [Figure 12] Figure 12 is a data flow diagram (DFD) showing an example of a data flow in accordance with the technology of this disclosure. [Modes for carrying out the invention]

[0023] The system disclosed herein addresses various performance issues faced by existing 4G / LTE antenna arrays when they are repurposed for providing 5G voice and data services. Passive structures (e.g., dielectric lenses) designed according to the technology of this disclosure improve signal coverage when incorporated into a phase-controlled antenna array. For example, passive structures designed according to the technology described herein may enable a phase-controlled antenna array to maintain its beamwidth without increasing it, even as the azimuth angle increases.

[0024] Figures 1A and 1B illustrate the difference between the enhanced signal coverage provided by a corresponding phase antenna array with a passive structure designed based on the technology of this disclosure and the signal coverage provided by an existing phase antenna array. Figure 1A shows the signal coverage provided by a three-sector antenna array with a currently available passive structure (e.g., dielectric lens). System 10A in Figure 1A provides three radio frequency (RF) beams positioned at the boresight angle of each sector antenna (i.e., 0 degrees from each antenna), spaced equally at 120 degrees apart, and reaching the cell boundary 12. Figure 1A shows a signal blind zone 14, which is a non-limiting example of a signal coverage gap (a so-called "dead spot"). In addition to the signal blind zone 14, system 10A also includes two other signal coverage gaps. The RF beams in the regions of these signal coverage gaps exhibit beamwidths that are spread proportionally to their respective azimuth angles. These widened beam widths reduce the beam's penetration depth, resulting in reduced signal coverage, or a dead zone, before it reaches the periphery represented by the cell boundary 12.

[0025] Figure 1B shows the signal coverage provided by a three-sectoral antenna array with a passive structure (e.g., dielectric lens) designed according to the art of this disclosure. All 16 RF beams of system 10B in Figure 1B reach the cell boundary 12, as their beam width does not increase with respect to the azimuth angle and their gain is uniform with respect to the azimuth angle. The radiation pattern of the boresite provides signal coverage over the entire 360-degree sweep of system 10B, which is shown by an unrestricted example of the signal coverage zone 16.

[0026] As illustrated in the example in Figure 1B, a passive structure (e.g., a dielectric lens) designed according to the techniques of this disclosure enables signal coverage across the entire cell site, even when retrofitted to existing three-sector antenna array infrastructure. The passive structure designed using the automated design techniques of this disclosure enables the cell-wide signal coverage provided by System 10B. In some cases, the performance of a passive structure designed according to the techniques of this disclosure can be verified by field testing in an integrated state with a phase antenna array. In other cases, the performance of a designed passive structure can be verified by simulation-based methods, such as modeling the design performance using simulation tools.

[0027] The following describes an example workflow based on the simulation-based verification technique of this disclosure. The system of this disclosure may execute a script agent to generate logical representations of various lens shapes. The simulation tool executed by the system of this disclosure (e.g., an electromagnetic field solver environment) may read one or more files (e.g., .stl files) for the purpose of performance modeling. As an example, the simulation tool may read one .stl file per simulation.

[0028] The simulation tool can perform automated electromagnetic field (EM) modeling under a simulation configuration in which the designed passive structure is placed on a phase antenna array. The system of this disclosure can collect modeled performance data in a structured format for each loaded passive structure design. The system of this disclosure can generate an index value (FoM) for the passive structure design based on the applicable performance metrics modeled by the simulation tool with respect to the phase antenna array in which the passive structure design is integrated.

[0029] The system of this disclosure may select the next sample (passive structural design) and iterate through the above workflow based on the FoM assigned to the passive structural design evaluated in the current pass, and (if available) the FoM based on past run results. By repeating the above workflow for multiple passive structural designs, the system of this disclosure may generate a dataset of actual modeling results for various lens shapes integrated into a phase antenna array (e.g., a 5G antenna array).

[0030] To achieve the above results in an efficient use of computing resources, the system of this disclosure incorporates Bayesian optimization techniques. Bayesian optimization can be described as a global optimization technique that attempts to find an optimal value within a limited number of function evaluations (i.e., "budget-constrained") for a potentially high-cost process that is non-convex and non-differentiable. In the context of the passive structural design generation techniques described herein, the system of this disclosure leverages Bayesian optimization techniques to reduce the resource costs associated with running EM simulations. Instead of brute-force searching the entire input space and running simulations at each input point, the system of this disclosure incorporates Bayesian optimization to reduce the number of simulations performed to obtain the optimal passive structural design.

[0031] Figure 2 is a block diagram showing an example implementation of the passive structure generator 20 of this disclosure. While Figure 2 shows one implementation of the passive structure generator 20 consistent with each aspect of this disclosure, it is understood that other architectures (whether single-device architecture or distributed architecture) are also consistent with each aspect of this disclosure in terms of the functionality of the passive structure generator 20 described in Figure 2. The passive structure generator 20 is described, in non-limiting examples, as a device that performs the functions of both a learning system and a user system. In other examples, a system consistent with this disclosure may be configured to function only as a learning system or only as a user system. Thus, it is understood that various configurations consistent with the technology of this disclosure may exist, and that Figure 2 shows a non-limiting example of a system configuration consistent with this disclosure.

[0032] In the example shown in Figure 2, the passive structure generating device 20 comprises a processing circuit 14 and a memory 16. In some examples, the processing circuit 14 and the memory 16 may be integrated into a single hardware unit, such as a system-on-a-chip (SoC) or integrated circuit (IC). The processing circuit 14 may represent one or more processors or processing units. This may include a multicore processor, a controller, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), processing circuits (e.g., fixed-function circuits, programmable circuits, or any combination thereof), or equivalent discrete logic circuits or integrated logic circuits. The memory 16 may be any form of memory for storing data and executable software instructions. Examples include random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), and flash memory. The memory 16 may be implemented to include one or more memory units.

[0033] The memory 16 and the processing circuit 14 are combined to provide a computing platform for running the operating system 22. The operating system 22 provides a multitasking environment for running one or more software components 30. As shown in the figure, the processing circuit 14 is connected to external systems and devices, such as devices via one or more communication networks, via an input / output (I / O) interface 18. The I / O interface 18 may include network interface hardware, such as wired and / or wireless network interface controllers (NICs), for communication via the communication link 24.

[0034] In the specific example in Figure 2, communication link 24 represents one or more network-enabled communication connections, such as a connection to one or more packet-switched networks collectively referred to as a "communication network" in the context of Figure 2. A communication network can be a data-enabled telephone network (e.g., a cellular data network), a wide-area network (e.g., the Internet), a public network (e.g., the Internet), a private network (e.g., a LAN or enterprise network), or any other network that enables data communication, or any two or more combinations of these networks.

[0035] Communication link 24 connects the passive structure generation device 20 to a communication network, enabling it to connect to other devices via that network. Communication link 24 may include wired connections (e.g., Ethernet® connections), wireless connections (e.g., Wi-Fi® connections), or a combination thereof. In Figure 2, the I / O interface 18 also mediates communication between the passive structure generation device 20 and one or more remote devices 26 via communication link 24. In the example in Figure 2, communication link 24 represents one or more local connections, such as personal area network (PAN) connections like near-field communication (NFC) pairing, Bluetooth® pairing, Zigbee® pairing, or connections via a local area network (LAN). Remote devices 26 may include computing devices (e.g., devices deployed in a dielectric lens manufacturer), additive manufacturing devices (so-called "3D printers"), etc.

[0036] In the implementation shown in Figure 2, bus 36 provides inter-component connectivity between the processing circuit 14, memory 16, and I / O interface 18. Bus 36 may be a half-duplex or full-duplex bus providing data transfer capability between the processing circuit 14, memory 16, I / O interface 18, and / or other hardware components. Bus 36 may represent a system bus or various types of computer buses and may include one or more bus networks. Regardless of the implemented topology, bus 36 may include various component connectivity hardware conforming to IEEE first to fourth generation bus or bus network technologies, or other standards to be defined in the future.

[0037] In the specific example shown in Figure 2, the software components 30 of the passive structure generation device 20 include a script agent 30A, an electromagnetic field (EM) simulator 30B, and a Bayesian optimizer 30C. In some examples, one or more of the software components 30 may represent executable software instructions in the form of a software application, software package, software library, hardware driver, application programming interface (API), etc. Furthermore, any of the software components 30 may operate to cause the passive structure generation device 20 to output or receive data via the I / O interface 18 during execution.

[0038] Of the memory 16, the portion providing non-volatile storage and / or long-term storage functions supports local storage of the data repository 34 in the passive structure generator 20. In the example in Figure 2, the data repository 34 includes the initial passive structure design 34A, performance metrics 34B, metric data (FoM) 34C, generated samples 34D, and optimized passive structure design 34E. In some examples, the software component 30 can perform read / write functions on the data repository 34, retrieve and use information from the data repository 34, or modify stored information. In implementations where the passive structure generator 20 is a distributed computing system, one or more of the data repositories 34 may be physically located remotely from the processing circuit 14, in which case the software component 30 can access the data repository 34 using the NIC hardware of the I / O interface 18.

[0039] One or more software components 30 may operate to activate processing circuits 14 and memory 16 to retrieve data from one or more data repositories 34 for various purposes, such as generating passive structural designs or optimizing the design generation process. The passive structural generation device 20 may train a learning model 32A or run a trained model 32B among the AI / ML models 32, in accordance with aspects of this disclosure. Further details will be provided later.

[0040] Figure 3 is a flowchart of an example workflow 40 of the present disclosure. To initiate an iteration of workflow 40, script agent 30A may run script agent to generate an initial passive structure design 32A (42). The initial passive structure design 32A includes logical representations of various dielectric lens shapes and / or form factors. In a particular iteration of workflow 40, EM simulator 32B may model the performance of a phase antenna array with lenses designed according to one of the initial passive structure designs 32A (44).

[0041] As a non-limiting example, the EM simulator 30B can take in a single .stl file containing the input data to be used in the current simulation path. As a result of running the EM simulation, the EM simulator 32B can generate a Figure of Merit (FoM) for the corresponding initial passive structure design 34A and store the generated FoM value in the FoM data 34C. Based on the FoM value stored in the FoM data 34C for the latest run of the EM simulator 30B, the acquisition function 58 is used as part of the Bayesian processing 30C to select a new design from within the search space. This is for an enhanced dielectric lens design (46).

[0042] The Bayesian optimization techniques implemented by the Bayesian optimizer 30C can be described as global optimization techniques that attempt to find optimal values ​​within a constrained number of function evaluations (also known as "budget limits") for high-cost processes that may be non-convex and non-differentiable. In the context of the passive structure design generation techniques described herein, the Bayesian optimizer 30C implements these techniques to reduce the resource costs associated with numerous runs of the EM simulator 30B. Instead of searching for the optimal dielectric lens shape by brute force across the entire input space and running the EM simulator 30B at every input point, the Bayesian optimizer 30C enables obtaining the optimal passive structure design while reducing the number of times the EM simulator 30B is executed.

[0043] Based on the recommendations in step 46, the Bayesian optimizer 30C may provide the updated passive structure design to the EM simulator 30B (feedback loop 48). By iteratively executing the process of the Bayesian optimizer 30C (46) and the feedback loop (48), the passive structure generator 20 can reduce the number of runs of the EM simulator 30B while maintaining the validity of the dielectric lens design modeled by the reduced runs of the EM simulator 30B. The Bayesian optimizer 30C also stores the optimized dielectric lens shape information in the optimized passive structure design 34E (38).

[0044] Non-limiting examples of performance metrics 34B include gain, beamwidth (which may be specified as "full width at half maximum" in some embodiments), and maximum gain (which is an indicator of beam depth). In various use cases consistent with each aspect of this disclosure, each data point included in performance metrics 34B may be associated with a specific scan angle at which the EM simulator 30B generated the performance metric.

[0045] Figure 4 is a flowchart illustrating an example process 50 that the passive structure generator 20 may implement to perform the Bayesian optimization technique of this disclosure. Process 50 may begin with a script agent 30A executing a script to generate a set of initial sample tuples (52). The structure of the tuple may combine one description of an initial passive structure design 34A with FoM values ​​(e.g., stored in FoM data 34C) generated based on a simulation run of the EM simulator 30B. The passive structure generator 20 includes an AI / ML model 32. The AI / ML model 32 includes a trained model 32A. The Bayesian optimizer 30C can fit the trained model 32A using Gaussian regression, where the initial passive structure design 34A obtained in the current run is used as the independent variable and the FoM values ​​of the corresponding FoM data 34B are used as the dependent variable. More specifically, the Bayesian optimizer 30C may train the trained model 32A using this Gaussian regression technique (54). This single execution (pass) of the learning process is represented as step 54.

[0046] The Bayesian optimizer 30C may generate a random set of sample parameter values ​​and provide these random sample parameter value sets to the trained model 32A (56). Alternatively, the Bayesian optimizer 30C may perform an optimization process on the trained model 32A to obtain a set of optimized samples. For example, the Bayesian optimizer 30C may generate various dielectric lens shapes based on the initial sample tuple evaluated in the current iteration of process 50 and store one or more representations of these generated shapes in the generated samples 34D. The Bayesian optimizer 30C may then evaluate the samples stored in the generated samples 34D by running an acquisition function (58) in the latest iteration of process 50. Running the acquisition function yields a predicted FoM value for each sample set in the generated samples 34D.

[0047] The acquisition function weights the sample value set based on both the likelihood of generating a higher FoM value and the uncertainty of the FoM prediction for each sample of the trained model 32A, thereby ranking the sample value set. The Bayesian optimizer 30C determines the next sample to evaluate by selecting the most weighted points or multiple points. Since further evaluation may be computationally intensive, the Bayesian optimizer 30C conserves computing resources by selecting samples from the generated samples 34D that are expected to produce the desired performance and also contribute to improving the performance of the trained model 32A.

[0048] The EM simulator 30B reads one or more of the generated samples 34D selected by the acquisition function. In some examples, one or more of the generated samples 34D may represent random samples or randomly generated samples. Each run of the EM simulator 30B constitutes an objective function. The EM simulator 30B evaluates the constrained search samples (i.e., selected from the generated samples 34D) based on the objective function and returns the simulation results in the form of FoM values. If the constrained samples are represented by the symbol μ, the EM simulator 30B performs an EM simulation using the μ samples and generates a μ tuple (62). The μ tuple contains the lens design represented by the μ samples and the FoM obtained by the EM simulator 30B evaluating the antenna performance of the phase antenna array augmented with the μ samples. The passive structure generator 20 may then provide the μ tuple to the learning model 32A as training data for subsequent learning passes (or relearning passes), thereby returning to step 54.

[0049] The passive structure generator 20 can further retrain the learned model 32A by iterating through steps 54 to 62. In some examples, the μ-tuples can be stored in one or more of the data repositories 34. For both the initial sample tuples and the μ-tuples, the Bayesian optimizer 30C may normalize all values ​​before performing Gaussian regression, although the original values ​​may be retained in the data repository 34. Examples of normalization ranges based on the Gaussian regression described above include the range from 0 to 1 or the range from -1 to 1. After a certain number of training and retraining iterations, the Bayesian optimizer 30C may determine that either design has achieved the desired performance or that the optimization has reached a predetermined number of iterations. The passive structure generator 20 outputs the trained model 32B (64).

[0050] In some examples, the passive structure generator 20 may invoke the I / O interface 18 to send the trained model 32B to an external device, such as a remote device 26, via the communication link 24 for deployment. In another example, the deployment unit 30C may invoke the I / O interface 18 to save a copy of the trained model 32B to a removable storage device 28. The removable storage device 28 may be any type of non-volatile storage medium, such as an external hard disk, solid-state drive (SSD), USB flash drive, or CD. In these examples, the passive structure generator 20 enables the other device to execute the trained model 32B in the execution phase to generate a passive structure design based on the Bayesian optimization technique of this disclosure.

[0051] In some further examples, the passive structure generator 20 may execute the pre-trained model 32B itself during the execution phase and generate an optimized passive structure design 34E using the computationally intensive Bayesian optimization technique of this disclosure. Alternatively, one or more samples of the optimized passive structure designs 34E may be used as training data for other models. Using one or more optimized passive structure designs 34E for training other models provides the technical advantage of improved data accuracy regarding the performance of the models for applications used outside of Bayesian techniques, and the technical advantage of resource saving by efficiently training these other models with fewer (more optimized) training samples.

[0052] Figure 5 is a graph 60 showing the clustering of sample sets within the constraints imposed according to the technology of this disclosure. The initial development of the platform described herein incorporated the use of various initial datasets. Regarding the definition of dimensions, the pitch and width of the ground plane and signal plane were set to be identical. In this specification, “pitch” means the distance from the start of one thread to the start of the next thread in the grid, and “width” means the linear dimension across the metal trace, measured in a manner similar to that used to measure the gauge of an electric wire. In this specification, “S11” means the signal traveling down the coplanar waveguide (CPW) (reflection coefficient, indicating how much power was reflected at the port), and “S21” means the propagating signal within the CPW (transmission coefficient, indicating how much power was lost during transmission).

[0053] For the sampling shown in Graph 60, the dimensions were constrained by two boundary conditions and a set of ranges for each dimension. The samples were distributed widely and artificially and randomly within the domain space. The space was divided into sections along the dimensions of the independent variable, and then randomly distributed throughout the regions specified in those sections. In other examples consistent with this disclosure, sampling may be performed in a similar manner, increasing the number of dimensions by adding additional relevant boundary conditions. The clustering 66 in Figure 5 shows a distribution of pitch and width constrained by transparency values. A minimum transparency of CPW is set, which is defined as the percentage of voids in the conductive mesh. This is given by equation (1) below:

number

[0054] Regarding the connectivity of the EM simulator 30B, dimensions were used to change the setting / running configuration, thereby programmatically building and running the model. The results were exported and referenced in the optimization process by the Bayesian optimizer 30C. The objective function represented by the simulation run by the EM simulator 30B is obtained by leveraging the model to calculate its value and importing the parameters of S11 and S21 to obtain the "actual" value. In this way, the Bayesian optimizer 30C is configured to optimize the objective function according to the techniques of this disclosure.

[0055] Aspects of the definition of individual FoMs in FoM data 34C are described herein. FoM values ​​are centered around the S11 and S21 parameters. The EM simulator 30B may combine the values ​​of the S11 and S21 parameters within a function. There are several different versions of the function that are consistent with each aspect of the EM simulator 30B and its functions described herein. In some examples, the EM simulator 30B may minimize the absolute value of the S11 parameter and maximize the value of the S21 parameter within the FoM. In one exemplary embodiment of the function, if the average of the absolute values ​​of the S11 parameter across all frequencies (denoted as "|S11|avg") exceeds a set minimum value (denoted as "S11min"), then the average of the absolute values ​​of the S21 parameter across all frequencies (denoted as "|S21|avg") is used to determine the FoM. In one example of the final function for FoM, the EM simulator 30B checks to exclude outliers that the absolute value of S11 at each frequency is greater than S11min, and the absolute value of S21 is greater than the set minimum value of S21 (denoted as "S21min"). Otherwise, the EM simulator 30B treats each parameter as its respective minimum value. Tables 1 and 2 below show various aspects of the FoM decision function described above. [Table 1] [Table 2]

[0056] The EM simulator 30B may average and combine the values ​​listed above according to the following equation (2):

number

[0057] Regarding the acquisition functions mentioned in Fig. 5, multiple different acquisition functions were evaluated. There are trade - offs in using these different functions. The Upper Confidence Bound (UCB) function shown below has a more exploitative nature rather than exploratory. It prioritizes short - term benefits by locally maximizing the reward at the current best position within the domain. On the other hand, the Expected Improvement (EI) function shown below sacrifices short - term rewards and instead pursues the possibility of better long - term rewards by exploring a wider area of the domain and leveraging knowledge acquisition. The Probability of Improvement function shown below has a property that lies between the exploitative nature of UCB and the exploratory nature of EI.

[0058] Probability of Improvement: x = argmax x , + , , + , + , , , + , E(u(x)|D), u(x)=0 (f < f + when), 1 (f ≥ f + when) PI(x)=P(f(x)≥μ + +ζ)⇒PI(x)=Φ(Z), Z=(μ(x)-μ + -ζ) / σ(x) Here: μ<000000​​​​​​​​​​​​​​​​​​​​​​​​​ UCB(x) = μ(x) + βσ(x) (for β>0, for maximization problems)

[0059] Figure 6A is a graph 70 showing the convergence of the FoM (FoM) over multiple runs of the EM simulator 30B for the EI function located on the exploratory side of the acquired function spectrum. The plotted line 72 shows the FoM data for the initial point, and the plotted line 68 shows the final convergence of the portion of the FoM data 34C corresponding to the selected point, obtained by simulation using a phase antenna array enhanced with one of the optimized passive structural designs 34E.

[0060] Figure 6B is a graph 74 showing the distribution of width and pitch for the EI acquisition function related to Figure 6A.

[0061] Figure 7 is Graph 80, which shows the convergence of FoM over multiple runs of the EM simulator 30B. In a specific use case in Graph 80, clear convergence (or stabilization) of FoM is shown in fewer than approximately 40 iterations, and this state is maintained over approximately 80 runs of the EM simulator 30B. The plot line for "initial points" shows the FoM data 34C calculated by the EM simulator 30B for the initial sample tuple mentioned in relation to Figure 4, while the plot line for "selection points" shows the FoM data 34C calculated by the EM simulator 30B for the μ tuple based on the search space constraints imposed by the Bayesian optimizer 30C.

[0062] Figure 8 is a graph 82 showing the convergence of FoM over multiple runs of the EM simulator 30B. As shown by the FoM plot line 84 in Figure 8, the maximum FoM increased in steps multiple times as the optimization continued after approximately 25 iterations, reaching convergence (or a general stable state) after approximately 33 iterations.

[0063] Figures 9A and 9B show various dielectric lens shapes generated by the passive structure generator 20 according to the Bayesian optimization design technique of this disclosure. In Figure 9A, the passive structure generator 20 generates dielectric lens design 86 using the following polynomial coefficient set: (1, 0, 1, 0, 1, 0.5, 0, 1). In Figure 9B, the passive structure generator 20 generates dielectric lens design 88 using the following polynomial coefficient set: (1, 1, 1, 1, 1). For both dielectric lens designs 86 and 88, the passive structure generator 20 may generate the dielectric lens shape using n-th order polynomial coefficients, which are two-dimensional representations, and generate the three-dimensional shape by rotating those polynomials 360 degrees. The passive structure generator 20 generates both dielectric lens designs 86 and 88 using the solid-of-revolution technique described above.

[0064] An example calculation formula that the EM simulator 30B may execute to calculate a predetermined data point in the FoM data 34C is shown in equation (3) below.

number

[0065] In some examples following the art of this disclosure, the EM simulator 30B may set the constant α to 0.7. As is evident from equation (3), the FoM of a phase antenna array not extended by any dielectric lens is zero (0). Therefore, any FoM value greater than zero (0) indicates an improvement over a phase antenna array without lenses. The weighting of the gain is based on the scanning angle and whether the gain of the phase antenna array extended by the dielectric lens exceeds the weighting for the gain of a phase antenna array not extended by any lens. If the gain of the lensed antenna is less than the gain of the phase antenna array configuration without lenses, a penalty of coefficient 5 (5 times) is imposed, regardless of the scanning angle. If the gain of the phase antenna array extended by the dielectric lens exceeds the gain of the phase antenna array configuration without lenses at small scanning angles, the weighting coefficient is 1 (1 times). In this scenario, if the scanning angle is large, the weighting coefficient is 5 (5 times).

[0066] In accordance with the specific non-restrictive examples in this disclosure, additional modifications to the calculations above may include one or more of the following: (i) Set the gain to minimum_no_lens_gain for the missing scan angle; (ii) Setting the beam width to the average beam width for missing values; (iii) For a case with a lens, interpolation along the scanning angle axis with a resolution of +2 degrees; and / or (iv) Subtract the gain variance or standard deviation (std) from the FoM value.

[0067] Therefore, the FoM decisions described above can be summarized according to the following five basic principles: (i) When there is lensing, the FoM increases as the gain is higher and the beam width narrows, especially at large scanning angles; (ii) If, at any scanning angle, the gain with lens extension is less than the gain without lens extension, the FoM receives a significant penalty (weighting coefficient of 5); (iii) If, at all scanning angles, the gain of the antenna with lens extension is higher than the gain of the antenna without lens, the gain at large scanning angles is given a higher weight than the gain at small scanning angles; (iv) In an antenna array with lenses, the gain improvement obtained compared to the case without lenses is weighted more highly than the reduction in beam width, with the former being weighted 70% and the latter 30%; (v) The smaller the variation of the gain with respect to the scanning angle, the larger the FoM. These points are shown in Table 3 below. [Table 3]

[0068] Figure 10 is a graph 90 showing one of the performance indicators 34B (i.e., peak gain) of various phase antenna arrays at various scanning angles. The control line 92 plots the peak gain of a phase antenna array without lens extension as the scanning angle increases. The conventional lens line 94 plots the peak gain of a phase antenna array extended with a conventional lens as the scanning angle increases. The automated design line 96 plots the peak gain of a phase antenna array extended with a dielectric lens designed by a passive structural design apparatus 20 using the technology of this disclosure as the scanning angle increases.

[0069] As shown by the comparison between the automated design line 96 and the conventional lens line 94 and control line 92, the phase antenna array extended with dielectric lenses designed according to the Bayesian-optimized automated design technique of this disclosure provides a larger and more consistent improvement in terms of peak gain (e.g., expressed as beam scanning range) compared to other scenarios. Thus, by implementing the technique of this disclosure, the passive structural design apparatus 20 leverages Bayesian optimization to reduce the consumption of computational resources for generating the dielectric lens designs while generating dielectric lens designs that improve performance index 34B (in this case, peak gain).

[0070] FoM data 34C shows an FoM value of 1.03 for the dielectric lens associated with the automated design line 96, while it shows an FoM value of -8.49 for the conventional dielectric lens associated with the conventional lens line 94. Furthermore, based on the phase antenna array extended by one or more optimized passive structural designs 34E, the FoM data 34C generated by the EM simulator 30B also shows significant improvement with respect to other types of data stored in the performance index 34B. For example, the phase antenna array extended with the dielectric lens having the above FoM value of 1.03 showed a significant reduction in beam width with increasing scan angle in simulations performed by the EM simulator 30B compared to the unextended or conventional dielectric lens extended phase antenna array.

[0071] Figure 11 shows various examples of the optimized passive structure design 34E and aspects of comparison with conventional passive structure designs. Dielectric lens designs 98A and 100A show two examples of shapes included in the optimized passive structure design 34E. Sectional views 98B and 100B show cross-sections of dielectric lens designs 98A and 100A, respectively. In contrast, conventional lens design 102A and sectional view 102B show renderings and cross-sectional shapes of conventional dielectric lens designs, respectively, and as shown in Figure 11, conventional lens design 102A and sectional view 102B reflect a hemispherical lens design.

[0072] The following is an overview of the simulation results for the design shown in dielectric lens design 98A and cross-sectional diagram 98B (collectively referred to as "automatic passive structure design 98"). For automatic passive structure design 98, the modeled dielectric constant is 1.49, the loss tangent is 0.0032, the diameter is 64 mm, and the maximum height is 23 mm. These parameters were also set identically in comparative experiments using lenses conforming to the conventional lens design 102A and cross-sectional diagram 102B (collectively referred to as "hemispherical lens 102"). In these simulations, the EM simulator 30B was run iteratively while increasing the scanning angle, and observations were made until the side lobes became larger than the main lobe when plotting the beam width data points of the performance index 34B. In control experiments without dielectric lenses, the side lobes appeared at 36 degrees. In the phase antenna array with the hemispherical lens 102 added, the side lobes appeared at 30 degrees. In the phase antenna array with the automatic passive structure design 98 added, the side lobes appeared at 36 degrees. Therefore, the EM simulator 30B demonstrated that the automated passive structure design 98 provides a significant improvement in terms of side lobe appearance for the hemispherical lens 102.

[0073] As described above, both the phase antenna array without dielectric lenses and the phase antenna array with the automated passive structure design 98 exhibit side lobes at the same scanning angle and provide the same baseline at specific data points of performance index 34B. Regarding peak gain (as shown in Figure 10), the automated passive structure design 98 provides an improvement over the phase antenna array without dielectric lenses. Furthermore, simulations performed with the EM simulator 30B showed that the automated passive structure design 98 exhibited a significant reduction in beamwidth compared to the phase antenna array without dielectric lenses in the scanning angle range of 30 to 36 degrees. Therefore, the experiment demonstrated that the passive structure design device 20 provides particularly advantageous improvements in the expansion and construction of 5G antenna arrays while generating a computationally efficient optimized passive structure design.

[0074] Figure 12 is a data flow diagram (DFD) 110 showing an example data flow in accordance with the disclosed technique. According to the DFD 110, a design 104 (e.g., one of the initial passive structural designs 34A) is input to an objective function 106 (which may represent a single execution pass of the EM simulator 30B). This allows the EM simulator 30B to model one or more performance results of the design 104 and reflect them in a performance metric 34B. The EM simulator 30B may also form a dataset 112 based on data points of FoM data 34C assigned to the design 104.

[0075] The Bayesian optimizer 30C is fed the dataset 112 and performance data 114 to the inverse surrogate model 116. The performance data 114 may represent more desirable or target performance data points that may be included in the ultimately optimized passive structure design 34E. The Bayesian optimizer 30C uses the combination of dataset 112 and performance data 114 as input. Depending on the various aspects of the disclosure, the inverse surrogate model 116 may conform to one or more data-driven models in machine learning, deep learning, or reinforcement learning. Examples of architectures to which the inverse surrogate model 116 may conform include logistic regression, Gaussian processes, random forests, support vector machines, neural networks (e.g., multilayer perceptrons (MLPs), convolutional neural networks (CNNs), generative adversarial networks (GANs), variational autoencoders (VAEs), etc.), and reinforcement learning models (e.g., deep Q networks (DQNs)). Alternatively, it may be any combination of these models. The inverse surrogate model 116 may output a selected model 118 for each iteration, and after convergence is reached, it may output the best candidate 122 to the optimized passive structural design 34E.

[0076] The Bayesian optimizer 30C may be configured to optimize the features of a coplanar waveguide (CPW) composed of a metal mesh conductor material to maximize the transmission of electromagnetic waves at 5G frequencies while satisfying optical transmission requirements. Through the optimization techniques of this disclosure, the Bayesian optimizer 30C solves several technical challenges, including: (i) the partial or complete simultaneous use of different variable types; (ii) evaluating the current learning status of a surrogate model; (iii) finding a global maximum value while mitigating or eliminating the risk of being trapped in a local optimum; and (iv) mitigating or eliminating the incomplete or insufficient nature of random sampling. In the experiment, the expected improvement (EI) cost function EI(x) shown in equation (4) was used, where a trade-off parameter ξ can be optionally added. Increasing the value of ξ causes the acquisition function (shown in Figure 4, 58) to weight the standard deviation more than the amount of improvement, and the optimization becomes more exploratory.

[0077]

number

[0078] In some experiments conducted according to the Bayesian optimization techniques of this disclosure, optimizing the grid of the signal plane and ground plane CPW grid to maximize the S11 and S21 values ​​was defined as the minimum viable product. The grid parameters identified for optimization were the offset of the grid from the plane edges, the rotation about the axis normal to the plane, the shape of the polygons combined to form the grid, the length of the sides of each polygon, and the width of the conductor wires forming the grid. While the applications shown herein are specific to transparent antenna development, it will be understood by those skilled in the art that the optimization techniques described herein are also applicable in the context of other types of simulation work.

[0079] As data is increasingly utilized in design and Modeling as a Service (MaaS) becomes more common, computer-assisted automated optimization is widely applicable to all types of product development. Some experiments following this technique utilized a robust, general-purpose dedicated repository designed to be applicable to parameter optimization problems related to CPW development and simulations of dielectric lenses for 5G applications. The methodology described herein is also suitable for hyperparameter tuning in machine learning (ML).

[0080] To define and combine types of variables, the Bayesian optimizer 30C is configured to optimize based on both qualitative or "categorical" variables and quantitative variables. Furthermore, for certain quantitative variables, the Bayesian optimizer 30C can be restricted to deterministic values ​​to exclude irrelevant variable ranges. The Bayesian optimizer 30C can perform classifications by type of parameter, and depending on the classification, it can achieve technical improvements such as saving computational resources by narrowing the search domain to reduce simulation time and setting up surrogate models to handle different types of data more appropriately. An example of variable classification used in experiments conducted in accordance with this disclosure is shown below. When classified as "continuously ranged," if a variable is quantitative and within a given range, the Bayesian optimizer 30C runs process 50 to explore all possibilities within that range.

[0081] In categorical classification, the Bayesian optimizer 30C can designate a variable as "categorical" if it is one of a limited number of qualitative labels. For such categorical variables, the Bayesian optimizer 30C pre-determines that the variable is discrete and does not have an inherent arithmetic order. Therefore, the Bayesian optimizer 30C can eliminate range-based processing of categorical variables and use encoding (e.g., one-hot encoding) that restricts the treatment of the variable as a value. In addition, the Bayesian optimizer 30C can use categorical variables as classifiers for dividing a domain into smaller subdomains for evaluation.

[0082] In the "discrete" classification, if the Bayesian optimizer 30C specifies a variable as discrete (for example, one of a predetermined set of possible values), it sequentially considers the possible values ​​for that variable instead of analyzing a large number of candidate values ​​within a given range. This allows the Bayesian optimizer 30C to restrict solution candidates based on the discrete specification, and, similar to categorical variables, to partition the domain using specific value sets, enabling better evaluation. In the "combined discrete" classification, the Bayesian optimizer 30C can specify this classification for discrete parameters where all values ​​within a range should be considered in the parameterized optimization (process 50). In such cases, the Bayesian optimizer 30C can identify optimal values ​​for all other parameters that maximize the FoM34C value, regardless of the value of the combined discrete variable. In this way, the Bayesian optimizer 30C can perform robust optimization even for parameters that are unknown or highly variable in real-world environments.

[0083] The Bayesian optimizer 30C may implement certain technical elements of this disclosure to determine whether the optimization process is stalled in a local region. In a certain use case, the optimization process may stall by concentrating on a local minimum within a domain. The techniques of this disclosure provide a technical benefit to the Bayesian optimizer 30C by enabling it to evaluate which part of the domain is currently being explored. According to these techniques, the Bayesian optimizer 30C can divide a domain into buckets, subdomains, or “domain spaces” by identifying combinations of categorical and discrete variables. In a certain example, the Bayesian optimizer 30C may also divide a domain using quantitative variables, classifying the quantitative variables into buckets defined by different ranges or sets of ranges. Within these smaller domain spaces, the Bayesian optimizer 30C can use two quantitative variables to construct axes on a plane and evaluate the degree of sample concentration within that plane. The density of samples in the domain space is referred to herein as “localization.” The Bayesian optimizer 30C can determine the relative degree of concentration by comparing the localizations across different domain spaces. Within a given domain space, the Bayesian optimizer 30C can construct a polygon, or "convex hull," enclosing the samples present in that space. The Bayesian optimizer 30C can then calculate the localization of that domain space by dividing the number of samples in the domain space by the area of ​​the convex hull. This calculation is shown by the following equation (Equation (5)):

number

[0084] The Bayesian optimizer 30C can use these localization values ​​to guide the optimization process to domain spaces that have not been explored sufficiently. For example, the Bayesian optimizer 30C can improve the acquisition function by linking these localization values ​​to the ξ parameter in the acquisition function. In the two different optimization processes, the changes in localization for each domain space are different. Both use random samples for the acquisition function, but in the optimization process in which the ξ parameter is incorporated into the acquisition function, the value shown in the corresponding FoM data 34C increases more steeply, and convergence (stabilization) is achieved earlier with fewer runs of the EM simulator 30B. The ξ parameter is linked to the localization in the domain space to which each sample belongs, and the incorporation of the ξ parameter brings advantages to a more exploratory acquisition function.

[0085] According to each aspect of this disclosure, the Bayesian optimizer 30C can monitor the training state of the surrogate model 116. As part of the overall optimization, the surrogate model 116 is continuously trained and retrained. To monitor the training progress, the Bayesian optimizer 30C evaluates how much the surrogate model 116 has changed between each iteration of the optimization. This technique is called the “predictive differences” method. According to the predictive differences method, the Bayesian optimizer 30C performs the following series of operations (repeated as necessary): 1. Generate a set of samples that are evenly distributed across the entire domain provided to the surrogate model 116. 2. The surrogate model 116 is executed on the aforementioned sample to obtain predictions. 3. In the next optimization iteration, the updated surrogate model 116 is run again on the same sample to obtain a prediction. 4. Evaluate the difference between the prediction obtained in the current iteration and the prediction obtained in the previous iteration.

[0086] These differences are expressed as the absolute percentage difference between the current predicted value and the predicted value in the previous iteration. By considering how the maximum, mean, and range of the prediction difference change between each iteration, the Bayesian optimizer 30C can determine the importance of the impact each known value has on the operation of the surrogate model 116. If the Bayesian optimizer 30C determines that there is no significant change in any of the prediction difference indicators over a certain number of iterations (e.g., a threshold number), it can determine that the surrogate model 116 has been sufficiently learned for the solution space, i.e., has reached a "saturated state". In the experiment, a period of significant change in the predicted values ​​between iterations was observed up to the 317th iteration. After reaching the 317th iteration, the mean of the prediction difference became relatively stable, and no significant changes were observed in subsequent optimization iterations. Therefore, in this experiment, it is determined that 317 iterations represent the starting point of saturation in the surrogate model 116.

[0087] Furthermore, the Bayesian optimizer 30C can monitor the impact of a particular sample on a specific aspect of the surrogate model 116, or the progress of training that aspect, by using the prediction difference in domain space units. The Bayesian optimizer 30C can use the techniques of this disclosure to optimize the acquisition function shown in Figure 4, 58. In some examples, the Bayesian optimizer 30C can achieve systematic sampling across the entire domain by leveraging Particle Swarm Optimization (PSO). With the implementation of PSO, the Bayesian optimizer 30C can treat randomly selected samples as "particles" and evaluate the position of each particle using the acquisition function. This makes it possible to identify the optimal position for each particle ("best individual particle position").

[0088] Next, the Bayesian optimizer 30C formulates a vector between the current positions of the particle group and the corresponding best-particle positions. For example, the vectors between the current positions of all particles and the best-particle positions can be combined by vector addition to generate a single composite vector. Then, the Bayesian optimizer 30C applies this composite vector to each particle to move it to a new position.

[0089] The Bayesian optimizer 30C can iteratively execute this process until one or more predetermined termination conditions are met. This iterative process essentially represents the process by which samples slide across the entire domain along the gradient defined by the acquisition function. This process covers a wider range of the domain and applies the acquisition function to the domain in a more complete manner.

[0090] As an alternative to PSO, another sampling technique that formed the basis for the experimental execution of the Bayesian optimizer 30C is called "pseudo-continuous" sampling. In this specification, "pseudo-continuous" sampling refers to a method in which a step size is defined for all quantitative range variables, and based on that step size, all values ​​of the variable are sampled within the domain. The Bayesian optimizer 30C can consider latent values ​​located between steps by adding a random shift (offset) between steps to move the sample position between iterations of the optimization operation. This enables more precise and homogeneous domain sampling. Experiments using pseudo-continuous sampling have shown that samples mainly remain within the same domain region, exhibiting a more exploitative rather than exploratory nature. On the other hand, by combining pseudo-continuous sampling with PSO and ξ localization, the Bayesian optimizer 30C can maintain the advantages of continuous sampling while finding other maxima across the entire domain in a more exploratory way.

[0091] As described above, a series of experiments conducted using the passive structure generator 20 employed various exploratory and exploitative optimization methods. A comparison of these methods is described here. The Bayesian optimizer 30C performed optimizations using different base methods with the same initial samples and range. In each iteration, the values ​​corresponding to the samples evaluated by the objective function, as well as the transition of the maximum value of FoM, were evaluated from the FoM data 34C. Three different optimization methods were compared. One was a baseline optimization without using PSO or ξ, another was optimization using PSO, and the third was optimization that linked the ξ value in the acquisition function to the localization of the domain to which each sample belongs. The differences between them are that the ξ-based optimization aims to improve the acquisition function itself, the optimization using PSO aims to improve the optimization method of the acquisition function itself, and the baseline optimization serves as a control experiment.

[0092] Baseline optimization, performed as a control experiment, reached a local maximum equivalent to the other methods after 365 iterations. On the other hand, the optimization method that linked locality to the ξ parameter reached a comparable global maximum in just 151 iterations, discovering numerous local maximums along the way. Optimization using PSO (Particle Swarm Optimization) reached a similarly high maximum FoM in just 15 iterations, but entered a stable state after reaching that point and never exceeded the FoM again. The PSO method explored the entire domain more extensively than the other two methods, but in the process, it did not discover any high local maximums like those found by the other methods.

[0093] In the detailed description of embodiments herein, specific embodiments in which the present invention may be carried out are described with reference to the drawings. These illustrated embodiments are not intended to exhaust all embodiments of the present invention. Other embodiments are also applicable, and it will be apparent to those skilled in the art that structural or logical modifications may be made without departing from the scope of the invention. Accordingly, the following detailed description should not be constrained, and the scope of the invention is defined by the appended claims.

[0094] Unless otherwise explicitly stated, all numerical values ​​(characteristic dimensions, quantities, physical properties, etc.) used herein and in the claims are understood to be modified in all contexts by the words “approximately,” “roughly,” or “substantially.” Accordingly, the numerical parameters described in the claims and the above specifications are to be interpreted as approximate values ​​that can be varied as desired by a person skilled in the art to obtain the target characteristics based on the technical guidance disclosed herein.

[0095] In this specification and the appended claims, the singular forms “a,” “an,” and “the” are understood to encompass embodiments that include multiple subjects, unless the context requires otherwise. Furthermore, in this specification and the appended claims, the word “or” is used to mean “and / or” unless otherwise specified in the context.

[0096] In some cases, certain actions or events in any of the methods described herein may be performed in a different order, added, merged, or omitted entirely (i.e., not all actions or events are essential for the implementation of the method). Furthermore, in some cases, such actions or events may be performed not sequentially, but in parallel by multithreading, interruption, or multiple processors.

[0097] The technologies described herein may be implemented in hardware, software, firmware, or any combination thereof, at least in part. For example, aspects of the technologies described may be implemented by one or more processors (e.g., microprocessors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), etc.), or equivalent integrated circuits or discrete logic circuits, or a combination thereof. The terms “processor” or “processing circuit” are construed to mean any of the aforementioned logic circuits, either alone or in combination with other logic circuits, and also encompass other equivalent circuits. A control unit including hardware may also implement one or more of the technologies described herein.

[0098] Such hardware, software, and firmware may be implemented within the same device or in separate devices to support the various operations and functions described herein. Furthermore, any of the described units, modules, or components may be implemented as one or more logic devices, integrated or individually. The descriptions of the functional configurations of modules and units are intended to highlight their respective functional aspects and do not necessarily imply that those modules or units must be implemented by separate hardware or software components. Rather, the functions associated with one or more modules or units may be performed by individual hardware or software components or integrated into common or individual hardware or software components.

[0099] The technologies described herein may also be embodied or encoded in a computer-readable medium, such as a computer-readable storage medium containing instructions. Instructions embedded in or encoded in a computer-readable storage medium can be executed by a programmable processor or other processor to perform the method. Computer-readable storage media may include random access memory (RAM), read-only memory (ROM), programmable ROM (PROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disks, CD-ROMs, floppy disks, cassettes, magnetic media, optical media, or other computer-readable media.

[0100] This specification describes various embodiments. These embodiments and other embodiments are included within the scope of the following claims.

Claims

1. A passive structure generating device, A non-temporary computer-readable storage medium in which instructions are stored, The system comprises at least one processor coupled to the at least one non-temporary computer-readable storage medium, wherein the at least one processor executes the instructions. Receiving a passive structure, Using simulation technology, an index is generated based on the passive structure and frequency data. Generate a pool of samples, Using the acquisition function, select the final sample from the pool of samples, The aforementioned final sample is provided to the simulation method to generate performance indicators, where the aforementioned final sample is associated with a new passive structure. A passive structure generating device that outputs the new passive structure to a user interface, an external device, or at least one of the at least one non-temporary computer-readable storage medium.

2. The passive structure generating apparatus according to claim 1, further comprising generating a pool of samples by executing a script agent.

3. The passive structure generating apparatus according to claim 2, wherein executing the script agent includes executing a script that outputs one or more modified examples generated from a conventional passive structure design.

4. The passive structure generating apparatus according to claim 3, wherein the conventional passive structure design includes a hemispherical dielectric lens design.

5. The passive structure generating apparatus according to claim 1, wherein the use of the simulation technique includes running an electromagnetic (EM) simulator configured to output one or more modeled performance metrics for an input passive structure design.

6. The passive structure generating apparatus according to claim 5, wherein the one or more performance indicators include one or more peak gain or beam width.

7. The passive structure generating apparatus according to claim 6, wherein the beam width is associated with a corresponding operating angle.

8. The passive structure generation apparatus according to claim 5, wherein the EM simulator is configured to take in a single file for each run in order to represent a dielectric lens design.

9. The passive structure generating apparatus according to claim 1, wherein the acquisition function includes a probability of improvement function.

10. The passive structure generating apparatus according to claim 1, wherein the acquisition function includes an Upper Confidence Bound (UCB) function.

11. The passive structure generating apparatus according to claim 10, wherein the UCB function implements an exploitative function.

12. The passive structure generating apparatus according to claim 11, wherein the acquisition function includes an expected improvement (EI) function.

13. The passive structure generating apparatus according to claim 12, wherein the EI function implements an explorative function.

14. The passive structure generating apparatus according to claim 13, wherein the EI function includes a trade-off parameter associated with localization values ​​in domain space.

15. Generating a pool of samples is The passive structure generating apparatus according to claim 14, comprising generating a pool of samples based on the localization value, wherein the pool of samples is generated in a search space within a domain space determined based on the localization value.

16. The passive structure generating apparatus according to claim 14, wherein the localization value is a value indicating a region surrounding the pool of samples.

17. Generating indicators Calculating localized values, Calculate the trade-off parameter value based on the localized value. Calculating predictive indicators based on surrogate functions, To estimate the predictive uncertainty of alternative models, To generate a ranking of candidate samples based on the aforementioned prediction indicator, the prediction uncertainty of the alternative model, the localization value, and the trade-off parameter, A passive structure generating apparatus according to claim 14, including the following:

18. The passive structure generating apparatus according to claim 17, wherein the acquisition function is configured to calculate a localization value based on the area of ​​the convex hull surrounding the pool of samples and the number of samples contained within that convex hull.

19. To calculate a new passive structure, To generate multiple new passive structures, To train an inverse substitution model based on the aforementioned multiple new passive structures, To provide a set of performance parameter values ​​to the inverse substitution model, To receive a new lens design from the aforementioned reverse substitution model, A passive structure generating apparatus according to claim 1, including the following:

20. The passive structure generating apparatus according to claim 19, wherein the inverse substitution model includes at least one of a logistic regression model, a Gaussian process model, a random forest model, a support vector machine model, a neural network, or a Deep-Q network.

21. The passive structure generation apparatus according to claim 20, wherein the neural network includes at least one of a multilayer perceptron network (MLP), a convolutional neural network (CNN), a generative adversarial network (GAN), or a variational autoencoder (VAE).

Images