Passive structural design for phased antenna array

Abstract

Systems and techniques for generating passive fabric designs are disclosed. A passive structure constructed according to the generated design improves the performance of the phased antenna array into which the passive structure is incorporated. An apparatus includes a non-transitory computer-readable storage medium encoded with instructions, and a processor coupled to the non-transitory computer-readable storage medium. The processor is configured to execute instructions to: receive a passive fabric design; generating a figure of merit based on the passive structural design and the frequency data using a simulation technique; calculating a new passive structure design based on the passive structure and the quality factor by using a Bayesian method; and output the new passive fabric to at least one of a user interface, an external device, or at least one non-transitory computer-readable storage medium.

Description

Technical Field

[0001] This disclosure relates to systems and techniques for designing passive structures to be used in phased antenna arrays. Summary of the Invention

[0002] As part of upgrading current mobile network infrastructure to provide 5G voice and data services, millimeter-wave (mmWave) phased array antennas are being installed on 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 array providing 120° (i.e., + / - 60°) azimuth coverage within a given cell. Combined, the three sector antennas provide 360° omnidirectional coverage around the site. Common 4G / LTE systems operate at frequencies below 6 GHz, which have lower propagation loss compared to mmWave frequencies typically at or above 20 GHz. To provide network coverage at the same range (distance from the RAN cell site) as common 4G / LTE systems, mmWave antennas must (i) be highly directive and (ii) have a tunable radiation pattern. Antenna arrays comprising multiple radiating elements provide these enhancements in the context of mmWave equipment. Antenna arrays typically use massive MIMO architectures to provide spatial diversity (one aspect by which a base station can use the same time-frequency resources to communicate with multiple devices in the same cell with the help of highly directional antennas), as in the case of 5G system specifications.

[0003] However, high directivity introduces one or more reductions. Providing high directivity requires many elements in each phased array, and the azimuth scan range of the entire antenna system is limited by the beam widening as the phased array moves further away from the optical axis (known as the "antenna line of sight"). Therefore, a highly directional mmWave phased array cannot provide 120° coverage over a wide scan angle without introducing significant gain degradation. This gain degradation results in reduced coverage at sector seams within the communication cell.

[0004] The technology disclosed herein relates to passive structures (e.g., dielectric lenses) for designing wider scanning ranges of mmWave phased antenna arrays. Potential advantages provided by the passive structure designs of this disclosure (e.g., dielectric lens designs) include achieving narrower beamwidths. For example, lower-order arrays (e.g., arrays with fewer antenna elements) incorporating the passive structure designs of this disclosure can provide resolution similar to higher-order arrays (e.g., arrays with more antenna elements).

[0005] In one example, the passive structure generation device includes at least one non-transitory computer-readable storage medium and a processor coupled to the at least one non-transitory computer-readable storage medium. Instructions are stored thereon on the at least one non-transitory computer-readable storage medium. The at least one processor is configured to execute the instructions to: receive a passive structure; generate a quality factor based on the passive structure and frequency data using simulation techniques; calculate a new passive structure based on the passive structure and the quality factor using a Bayesian method; and output the new passive structure to at least one of a user interface, an external device, or the at least one non-transitory computer-readable storage medium.

[0006] In another example, a device includes interface hardware, memory hardware, and processing circuitry communicatively coupled to the memory hardware. The interface hardware is configured to receive a passive structural design. The memory hardware is configured to store the passive structural design received via the interface hardware. The processing circuitry is configured to generate a quality factor using analog techniques based on the passive structural design stored in the memory hardware and based on frequency data. The processing circuitry is further configured to calculate a new passive structural design using a Bayesian method based on the passive structural design stored in the memory hardware and the quality factor. The processing circuitry is configured to perform at least one of the following: outputting the new passive structural design via the interface hardware, and / or storing the new passive structural design in the memory hardware.

[0007] In another example, a method includes: receiving a passive structural design by processing circuitry of a computing device; and generating a quality factor by the processing circuitry based on the passive structural design and frequency data using analog techniques. The method further includes calculating a new passive structural design by the processing circuitry system using a Bayesian method based on the passive structural design and the quality factor, and outputting the new passive structural design by the processing circuitry system to at least one of a user interface, an external device, or at least one non-transitory computer-readable storage medium.

[0008] In another example, an apparatus includes components for receiving a passive structural design and for generating a quality factor based on the passive structural design and frequency data using analog techniques. The apparatus also includes components for calculating a new passive structural design based on the passive structure and the quality factor using a Bayesian method, and components for outputting the new passive structural design to at least one of a user interface, an external device, or a non-transitory computer-readable storage medium.

[0009] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed by processing circuitry of a computing device, the instructions cause the processing circuitry to: receive a passive structure design; generate a quality factor based on the passive structure design and frequency data using analog techniques; calculate a new passive structure based on the passive structure design and the quality factor using a Bayesian method; and output the new passive structure design to at least one of a user interface, an external device, or the at least one non-transitory computer-readable storage medium.

[0010] The passive structure design techniques disclosed herein offer several improvements in the field of phased antenna array design. In this way, the passive structure design of this disclosure reduces the design complexity and cost of phased antenna array systems by reducing the number of one or more of the phase shifters, amplifiers, and / or impedance matching networks required in the system. The passive structures designed according to the techniques of this disclosure provide these performance enhancements in the context of a three-sector antenna implementation, but in many cases, the infrastructure can be reduced to one or two antenna arrays, particularly in use cases covering smaller, denser areas of equipment deployment, such as urban city center areas. Furthermore, the improvements provided by the passive structures designed according to the techniques of this disclosure enhance cell-level performance and capacity, thereby potentially reducing the number of arrays required from a higher (e.g., system-level) perspective. Attached Figure Description

[0011] Figure 1A and Figure 1B The difference between signal coverage provided by a conventional phased antenna array and enhanced signal coverage of a corresponding phased antenna array equipped with a passive structure designed according to the present disclosure is illustrated.

[0012] Figure 2 This is a block diagram illustrating an example specific implementation of the passive structure generation device of this disclosure.

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

[0014] Figure 4 This is an example Figure 2 The passive structure generation device can realize a flowchart of an example process for performing the Bayesian optimization technique disclosed herein.

[0015] Figure 5 This is a graph illustrating the clustering of sample sets within the constraints imposed by the technology according to this disclosure.

[0016] Figure 6A This is a graph showing the convergence of the quality factor (FoM) of the expected improvement (EI) function at the exploratory end of the acquired function spectrum across multiple runs of the electromagnetic simulator (EM).

[0017] Figure 6B It shows that it is aimed at and Figure 6A A graph showing the width and spacing distribution of the associated EI acquisition functions.

[0018] Figure 7 This is a graph showing the convergence of FoM across multiple runs of the EM simulator.

[0019] Figure 8 This is a graph showing the convergence of FoM across multiple runs of the EM simulator.

[0020] Figure 9A and Figure 9B An example is illustrated by the Bayesian optimization design technique according to this disclosure. Figure 2 Various dielectric lens shapes generated by passive structure generation equipment.

[0021] Figure 10 It is a graph showing one of the performance metrics (i.e., peak gain) of various phased antenna arrays at various scanning angles.

[0022] Figure 11 Various aspects of the optimized passive structural design of this disclosure are illustrated compared to conventional passive structural designs.

[0023] Figure 12 This is a data flow diagram (DFD) illustrating an example data flow according to the technology of this disclosure. Detailed Implementation

[0024] The system disclosed herein addresses various performance deficiencies of existing 4G / LTE antenna arrays when repurposed for 5G voice and data service delivery. Passive structures (e.g., dielectric lenses) designed according to the techniques of this disclosure improve signal coverage when incorporated into phased antenna arrays. For example, passive structures designed according to the techniques described herein enable phased antenna arrays to maintain a non-increasing beamwidth even with increasing azimuth angle.

[0025] Figure 1A and Figure 1B The difference between signal coverage provided by a conventional phased antenna array and enhanced signal coverage of a corresponding phased antenna array equipped with a passive structure designed according to the present disclosure is illustrated. Figure 1A The example illustrates signal coverage provided by a three-sector antenna array equipped with currently available passive structures (e.g., dielectric lenses). Figure 1A System 10A provides three instances of radio frequency (RF) beams that reach the cell boundary 12 at line-of-sight angles of 120 degrees apart (i.e., zero degrees with respect to the corresponding antennas of the three-sector array). Figure 1ASignal dead zone 14 is illustrated, which is a non-limiting example of a signal coverage gap (or so-called "dead spot"). In addition to signal dead zone 14, system 10A includes two other signal coverage gaps. The radio frequency (RF) beams in these regions of the signal coverage gaps exhibit a beamwidth that is widened proportionally to the corresponding azimuth angle. The wider beamwidth associated with these RF beams reduces the corresponding beam depth, thereby causing a reduction in signal coverage or a dead zone before reaching the perimeter represented by cell boundary 12.

[0026] Figure 1B The example illustrates signal coverage provided by a three-sector antenna array equipped with a passive structure (e.g., a dielectric lens) designed according to the technology of this disclosure. Figure 1B All sixteen RF beams of system 10B reach the cell boundary 12 because the beamwidth of the beams does not increase with azimuth and the gain is uniform with azimuth. The radiation pattern along the line of sight provides signal coverage throughout the entire 360-degree scan of system 10B, as shown in a non-limiting example through signal coverage area 16.

[0027] If passed Figure 1B Examples illustrate how passive structures (e.g., dielectric lenses) designed according to the techniques of this disclosure achieve full-cell site signal coverage, even when retrofitted into an existing three-sector antenna array infrastructure. The full-cell signal coverage provided by System 10B is achieved through a passive structure designed using the automated design techniques of this disclosure. In some examples, the performance of the passive structure designed according to the techniques of this disclosure can be verified by field testing through integration into a phased antenna array. In other examples, the performance of the passive structure designed according to the techniques of this disclosure can be verified via simulation-based techniques, such as modeling the performance of the design using simulation tools.

[0028] The following describes an example workflow for a simulation-based verification technique according to this disclosure. The system of this disclosure can execute script agents to generate logical representations of various lens shapes. Simulation tools (e.g., electromagnetic solver environments) run by the system of this disclosure can ingest one or more files (e.g., one or more .stl files) for performance modeling purposes. In a non-limiting example, the simulation tool can ingest a single .stl file per simulation pass.

[0029] The simulation tool can perform automated electromagnetic (EM) modeling runs in a simulation configuration where the designed passive structure is placed on a phased antenna array. The system disclosed herein can collect performance data for modeling the passive structure design for each acquisition in a structured format. The system disclosed herein can generate a factor of quality (FoM) value for a given passive structure design based on the corresponding performance metrics modeled by the simulation tool for the phased antenna array integrated into the passive structure design.

[0030] Based on the FoM of the passive structural design assigned to be evaluated in the current iteration and (if available) the FoM results from previous runs, the system of this disclosure can select the next sample (passive structural design) for which to run the simulation and can iterate the above workflow. By iterating the above workflow for multiple passive structural designs, when integrated into a phased antenna array (such as a 5G antenna array), the system of this disclosure can generate a dataset of realistic modeling results for different lens shapes.

[0031] To achieve the aforementioned results in a computationally resource-efficient manner, the system disclosed herein incorporates Bayesian optimization techniques. Bayesian optimization can be described as a global optimization technique that attempts to find the optimal value of a non-convex, potentially non-differentiable, and expensive process within a constrained number of function evaluations (also known as a “finite budget”). In the context of the passive structure design generation technique described herein, the system disclosed herein utilizes Bayesian optimization techniques to reduce the resource costs associated with EM simulation runs. The system disclosed herein is configured to incorporate Bayesian optimization to reduce the number of simulation iterations performed to obtain the optimal passive structure design, rather than using an exhaustive search of the entire input space to search for the optimal passive structure design when running simulation iterations at each input point.

[0032] Figure 2 This is a block diagram illustrating an example specific implementation of the passive structure generation device 20 of this disclosure. Although Figure 2 A specific embodiment of the passive structure generation device 20 consistent with various aspects of this disclosure is shown; however, it should be understood that other architectures (whether monolithic or distributed) relative to the functionality described for the passive structure generation device 20 are also consistent with various aspects of this disclosure. As a non-limiting example, the passive structure generation device 20 is described as functionally performing both a training system and a exploitation system. In other examples consistent with this disclosure, the system may be configured to perform exclusively as a training system or exclusively as a exploitation system. Therefore, it should be understood that different configurations are consistent with the technology of this disclosure, and Figure 2 A non-limiting example of a system configuration consistent with this disclosure is illustrated.

[0033] exist Figure 2In the examples, the passive structure generation device 20 includes processing circuitry 14 and memory 16. In some examples, processing circuitry 14 and memory 16 may be integrated into a single hardware unit such as a system-on-a-chip (SoC) or integrated circuit (IC). Processing circuitry 14 may represent one or more processors or processing units, each of which may include one or more of a multi-core processor, controller, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), processing circuitry (e.g., fixed-function circuitry, programmable circuitry, or any combination of fixed-function circuitry and programmable circuitry), or equivalent discrete logic circuitry or integrated logic circuitry. Memory 16 may include any form of memory for storing data and executable software instructions, such as 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. In various specific implementations, memory 16 may include a single memory cell or multiple memory cells.

[0034] The combination of memory 16 and processing circuitry 14 provides a computing platform for executing operating system 22. Operating system 22 provides a multitasking operating environment for executing one or more software components 30. As shown, processing circuitry 14 is connected to external systems and devices, for example, via input / output (I / O) interface 18, such as via one or more communication networks. I / O interface 18 may incorporate network interface hardware, such as one or more wired and / or wireless network interface controllers (NICs) for communication via communication link 24.

[0035] exist Figure 2 In a specific example, communication link 24 represents one or more network-enabled communication connections, such as to... Figure 2 In the context of this document, a link of one or more packet-switched networks is collectively referred to as a “communication network”. A communication network may mean any of the following: a data-enabled telephone network (such as a cellular data network), a wide area network (such as the Internet), a public network (such as the Internet), a private network such as a local area network (LAN) and / or an enterprise network, or any other type of network that enables data communication, or any combination of any two or more of the networks listed above.

[0036] Communication link 24 communicatively couples the passive structure generation device 20 to a communication network, and via the communication network to other devices. Each of the communication links 24 may include one or more wired connections (e.g., Ethernet). ® (Connectivity), wireless connectivity (e.g., Wi-Fi) ™ (Connection) or a combination of wired and wireless communication connections. Figure 2In the illustrated example, I / O interface 18 also facilitates communication between the passive structure generation device 20 and one or more remote devices 26 via communication link 24. Figure 2 In a specific example, communication link 24 represents one or more local connections, such as connections to remote device 26 via a local area network (LAN) and / or personal area network (PAN), such as near field communication (NFC) pairing, Bluetooth, etc. ® Pairing, Zigbee ® One or more of the following: pairing, etc. Remote device 26 may include one or more of the following: computing devices (e.g., devices deployed at the dielectric lens manufacturing entity), additive manufacturing devices (e.g., so-called "3D printers"), etc.

[0037] exist Figure 2 In the specific implementation shown, bus 36 provides inter-component connectivity between processing circuitry 14, memory 16, and I / O interface 18. Bus 36 may represent a half-duplex or full-duplex bus providing data transfer capability between two or more of the processing circuitry 14, memory 16, I / O interface 18, and / or any other hardware component of the passive structure generation device 20. Bus 36 may represent various types of system buses or computer buses, including one or more bus networks. Regardless of the topology of the implementation, in various examples, bus 36 may incorporate various types of inter-component connectivity hardware, such as hardware conforming to any of the first, second, third, or fourth generation bus or bus network technologies as described by the Institute of Electrical and Electronics Engineers (IEEE) or other bus or bus network technologies defined in standards under development or to be adopted later.

[0038] exist Figure 2 In a specific example, the software component 30 of the passive structure generation device 20 includes a script agent 30A, an electromagnetic (EM) simulator 30B, and a Bayesian optimizer 30C. In some exemplary methods, one or more software components in software component 30 represent executable software instructions, which may take the form of one or more software applications, software packages, software libraries, hardware drivers, and / or application programming interfaces (APIs). Furthermore, any software component in software component 30, when executed, causes the passive structure generation device 20 to output and / or receive data via I / O interface 18.

[0039] The provision of non-volatile and / or long-term storage by memory 16 supports local storage of data storage library 34 at the passive structure generation device 20. Figure 2In the example, data store 34 includes an initial passive architecture design 34A, performance metrics 34B, quality factor (FoM) data 34C, generated samples 34D, and an optimized passive architecture design 34E. In some examples, software component 30 can implement read / write capabilities regarding data store 34, such as accessing and using information available from data store 34 and / or modifying information currently stored in data store 34. In the embodiment where passive architecture generation device 20 represents a distributed computing system, one or more data stores in data store 34 may be located partially or completely at a remote location from processing circuitry 14, and in these embodiments, software component 30 can access data store 34 using NIC hardware of I / O interface 18.

[0040] One or more software components in software component 30 can invoke processing circuitry 14 and memory 16 to access one or more data repositories in data repository 34 to retrieve information for various purposes, such as for passive structure design generation, optimization of the design generation process, etc. According to various aspects of this disclosure, passive structure generation device 20 can train a training model 32A of AI / ML model 32 and / or execute a trained model 32B of the AI / ML model, as described in more detail below.

[0041] Figure 3 This is a flowchart illustrating an example workflow 40 of this disclosure. To begin an iteration of workflow 40, script agent 30A can run a script agent to generate an initial passive structural design 32A (42). The initial passive structural design 32A includes logical representations of various dielectric lens shapes and / or shape factors. In a given iteration of workflow 40, EM simulator 32B can model the performance of a corresponding lens-enhanced phased antenna array designed based on one of the initial passive structural designs 32A (44).

[0042] In a non-limiting example, EM simulator 30B can ingest a single .stl file containing input data for the current simulation pass. As output of the EM simulation run, EM simulator 32B can generate a quality factor (FoM) value for the corresponding initial passive structural design 34A and store the generated FoM value in FoM data 34C. Based on the FoM value of the most recently run for EM simulator 30B stored in FoM data 34C, a new design is selected from the search space using acquisition function 58 as part of Bayesian process 30C to achieve an enhanced dielectric lens design (46).

[0043] The Bayesian optimization techniques implemented by the Bayesian optimizer 30C can be described as global optimization techniques that attempt to find the optimal value of a non-convex, potentially non-differentiable, and expensive process within a constrained number of function evaluations (also known as a “finite budget”). 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 the potentially large number of iterations performed by the EM simulator 30B. Instead of having the passive structure generation device 20 use an exhaustive search across the entire input space to search for the optimal dielectric lens shape (which would require running the EM simulator 30B for every input point in the input space), the Bayesian optimizer 30C reduces the number of runs of the EM simulator 30B required to obtain the optimal passive structure design.

[0044] Based on the recommendation obtained at point 46, the Bayesian optimizer 30C can provide an updated passive structure design to the EM simulator 30B (feedback loop 48). By iterating the Bayesian optimizer 30C run at points 46 and feedback loop 48, the passive structure generation device 20 reduces the number of runs of the EM simulator 30B while maintaining the relevance of the dielectric lens design modeled by these 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).

[0045] Non-limiting examples of performance metric 34B include values ​​such as gain, beamwidth (which in some instances may be specified as "half-power beamwidth"), and maximum gain (which is an indicator of beam depth). In various use case scenarios consistent with aspects of this disclosure, each data point included in performance metric 34B may be matched with a specific scan angle for which the EM simulator 30B generates performance metric 34B.

[0046] Figure 4 This is a flowchart illustrating an example process 50 that the passive structure generation device 20 can implement to perform the Bayesian optimization techniques of this disclosure. Process 50 may begin with a script agent 30A running a script to generate a set of initial sample tuples (52). The structure of the tuples may include a description of an initial passive structure design 34A that combines with FoM values ​​generated by a simulation run based on the EM simulator 30B (e.g., as stored in FoM data 34C).

[0047] The passive structure generation device 20 includes an AI / ML model 32. The AI / ML model includes a training model 32A. A Bayesian optimizer 30C can fit the training model 32A using Gaussian regression, using a tuple representing the initial passive structure design 34A of the current run as the independent variable and the corresponding FoM value of the FoM data 34B as the dependent variable. More specifically, the Bayesian optimizer 30C can use this Gaussian regression technique to train the training model 32A (54). A single pass of the training process for the training model 32A is represented at 54.

[0048] The Bayesian optimizer 30C can generate a set of random sample parameter values ​​and provide the random sample parameter set to the training model 32A (56). The Bayesian optimizer 30C can also run optimization on the training model 32A to obtain a set of optimized samples. For example, the Bayesian optimizer 30C can generate various dielectric lens shapes based on the initial sample tuple evaluated in the current iteration of process 50 and store representations of one or more of these generated shapes in the generated samples 34D. The Bayesian optimizer 30C can then evaluate the generated samples stored in the generated samples 34D in the most recent iteration of process 50 by running the acquisition function (58). By running the acquisition function at 58 on selected samples in the generated samples 34D, the Bayesian optimizer 30C produces predicted FoM values ​​for each corresponding set of generated samples 34D.

[0049] The acquisition function ranks the set of values ​​by weighting the probability of the set of values ​​generating a high FoM value and the uncertainty of the corresponding FoM value of the training model 32A relative to each sample. The Bayesian optimizer 30C selects the highest-weighted point, thus choosing the next sample to evaluate. Since further evaluation could represent a computationally intensive operation, the Bayesian optimizer 30C conserves computational resources by selecting samples from the set represented by the generated samples 34D that are expected to both have the desired performance and improve the performance of the training model 32A.

[0050] EM simulator 30B ingests generated samples from the generated samples 34D selected by the acquisition function. In some instances, one or more generated samples in the generated samples 34D may represent random samples or randomly generated samples. Each run of EM simulator 30B represents an objective function. EM simulator 30B uses the objective function to estimate the search constraint samples of the generated samples 34D, thereby returning the simulation results in the form of FoM values. If the search constraint samples are identified by the symbol μ, then EM simulator 30B uses μ samples to run the EM simulation to generate μ tuples (62). The μ tuples include the lens design represented by the μ samples and the FoM generated by EM simulator 30B when evaluating the antenna performance for a phased antenna array enhanced using μ samples. The passive structure generation device 20 can then provide the μ tuples to the training model 32A as training data for subsequent training iterations (or retraining iterations), thus returning to 54.

[0051] The passive structure generation device 20 can iterate 54 to 62 times to further retrain the trained model 32A. In some examples, the passive structure generation device 20 can store the μ-tuples into one or more data stores in the data store 34. The Bayesian optimizer 30C can normalize all values ​​before running them through Gaussian regression, relative to the initial sample tuples and the μ-tuples, but can store the original values ​​in the data store 34. The normalization range according to the Gaussian regression above includes the range of zero to one and / or the range of negative one to one, as well as other possible ranges. After multiple training and retraining iterations, the Bayesian optimizer 30C can determine that the desired performance of a design has been achieved, or that the optimization has run a set number of iterations. The passive structure generation device 20 can output the trained model 32B (64).

[0052] In some examples, the passive structure generation device 20 may invoke I / O interface 18 to deploy the trained model 32B to an external device, such as by signaling the trained model 32B to a remote device 26 via communication link 24. In some examples, the deployment unit 30C may invoke 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 represent any type of non-volatile storage medium, such as an external hard disk drive or solid-state drive (SSD), a USB flash drive, a CD, etc. In these examples, the passive structure generation device 20 enables other devices to run the trained model 32B during the execution phase to generate a passive structure design according to the Bayesian optimization techniques of this disclosure.

[0053] In some examples, the passive structure generation device 20 can run a trained model 32B during its execution phase to generate an optimized passive structure design 34E, while utilizing the computationally resource-saving Bayesian optimization techniques of this disclosure. In some examples, one or more samples of the optimized passive structure design 34E are provided as training data to train other models. Using one or more optimized passive structure designs from the optimized passive structure design 34E to train other models provides the technical advantage of data accuracy relative to the performance of models suitable for use in applications other than Bayesian techniques, as well as the technical advantage of resource conservation in training these other models more efficiently with a smaller (more optimized) training sample set.

[0054] Figure 5 This is a graph 60 illustrating a cluster of sample sets within the constraints imposed by the technology according to this disclosure. The initial development of the platform described herein incorporated the use of various initial datasets. In terms of dimensional definitions, the spacing and width of the ground plane and signal plane are set to be identical. As used herein, “spacing” refers to the distance between the start of one thread and the next thread in the mesh, while “width” refers to a linear measurement of the cross-metal trace in a manner similar to the specifications of a measuring conductor. As used herein, “S11” refers to the signal traveling along the coplanar waveguide or “CPW” (reflection coefficient, indicating how much power is reflected at the port), and “S21” refers to the reflected signal in the CPW (transmission coefficient, indicating how much power is lost during transmission).

[0055] In the sampling described in Figure 60, the dimensions are constrained by two boundary conditions and a set of ranges for each dimension. The samples are artificially and randomly distributed widely in the domain space. The space is segmented along the independent variable dimensions and then randomized within the regions specified by the segments. In other examples consistent with this disclosure, sampling can be performed in a similar manner, while adding further dimensions by adding additional associated boundary conditions. Figure 5 Cluster 66 shows the spacing and width distribution limited by transparency values. The minimum transparency of the CPW is established, defined as the percentage of empty space in the conductive mesh, as shown in the following formula (1):

[0056] …(1)

[0058] Compared to the connectivity of the EM simulator 30B, dimensions are used to change the setup / run configuration so that the model can be subsequently built and executed programmatically. The results are exported and then referenced by the Bayesian optimizer 30C during optimization. The objective function represented by the simulation run by the EM simulator 30B utilizes the model to compute values ​​and imports S11 and S21 parameters to achieve "actual" values. Therefore, the Bayesian optimizer 30C is configured to optimize the objective function according to the techniques of this disclosure.

[0059] This document describes various aspects of the definition of the various FoMs in FoM data 34C. FoM values ​​are centered around the S11 and S21 parameters. The EM simulator 30B can combine the S11 and S21 parameter values ​​in a function. Several different versions of this function are consistent with aspects of the EM simulator 30B and its functionality as described herein. In some examples, the EM simulator 30B can minimize the magnitude of the S11 parameter and maximize the value of the S21 parameter within the FoM. In one example implementation of this function, if the average of the S11 parameter amplitudes across all frequencies (denoted as "|S11|avg") is above a set minimum value (denoted as "S11min"), then the average of the S21 parameter magnitudes across all frequencies (denoted as "|S21|avg") is used as the FoM determination. In the example of the final function for FoM, the EM simulator 30B excludes outliers by checking that the S11 value for each frequency is greater than S11min and the S21 value for each frequency is above the set minimum value of S21 (denoted as "S21min"). Otherwise, the EM simulator 30B treats them as their respective minimum values. Tables 1 and 2 below illustrate various aspects of the FoM determination function described above.

[0060]

[0061] Table 1

[0062] <![CDATA[S21 min ]]> S21 <![CDATA[ ≥ |S21| avg： ]]> <![CDATA[|S21| avg ]]> <![CDATA[ ＜ |S21| avg： ]]> <![CDATA[S21 min ]]>

[0063] Table 2

[0064] The EM simulator 30B can average and combine the values ​​listed above according to the following formula (2):

[0065] …(2)

[0067] Compared to Figure 5The acquisition function referred to herein evaluates several different acquisition functions. There are trade-offs when using different functions. The Confidence Upper Limit (UCB) function shown below more accurately reflects the exploitation of privilege and maximizes the reward in a local region to the current best position in the domain in the short run. The Expected Improvement (EI) function shown below has the potential to sacrifice short-term rewards for better long-term rewards by further exploring a wider domain and utilizing knowledge gains. The improved probability function shown below lies somewhere between the exploitative nature of UCB and the exploratory nature of EI.

[0068] Probability of improvement:

[0069]

[0070] Where μ + It is the maximum average value, ζ = le-9, Φ is the Gaussian CDF, and It's a Gaussian PDF.

[0071] Note that f(x) is approximated by a surrogate Gaussian process that returns μ and σ for each x.

[0072] Desired improvements:

[0073]

[0074] Confidence ceiling:

[0075] Where β > 0 refers to the maximization problem.

[0076] Figure 6A Graph 70 shows the convergence of the FoM (Form of Mean) across multiple runs of the EI function at the exploratory end of the acquired function spectrum using EM simulator 30B. Plot line 72 shows the FoM data for an initial point, while plot line 68 shows the final convergence for a portion of the FoM data 34C applicable to selected points of the simulation representation run using an optimized passive structure design enhanced by an optimized passive structure design 34E.

[0077] Figure 6B It shows that it is aimed at and Figure 6A Figure 74 shows the distribution curves of the width and spacing of the associated EI acquisition functions.

[0078] Figure 7 This is graph 80, showing the convergence of FoM across multiple runs of the EM simulator 30B. The specific use case scenario of graph 80 shows the decisive convergence (or stabilization) of FoM within just under 40 iterations, and its retention in place until approximately 80 iterations of the EM simulator 30B's execution run. The line drawn relative to the "initial point" refers to the EM simulator 30B's convergence relative to the above. Figure 4 The FoM data 34C described is computed from the initial sample tuples, while the drawn line for the “selected point” refers to the FoM data 34C computed by the EM simulator 30B for the μ tuples based on the search space constraints imposed by the Bayesian optimizer 30C.

[0079] Figure 8 Graph 82 shows the convergence of FoM across multiple runs of the EM simulator 30B. (As shown by...) Figure 8 The FoM plot line 84 shows that as optimization continues after about 25 iterations, the maximum FoM step increases multiple times and converges (or reaches a roughly stable state) after about 33 iterations.

[0080] Figure 9A and Figure 9B Various dielectric lens shapes generated by the passive structure generation device 20 according to the Bayesian optimization design technique of this disclosure are illustrated. The passive structure generation device 20 uses the following polynomial coefficient set: (1, 0, 1, 0, 1, 0.5, 0, 1) to generate... Figure 9A The dielectric lens design 86. The passive structure generation device 20 uses the following set of polynomial coefficients to generate... Figure 9B The dielectric lens design 88 is (1, 1, 1, 1, 1). Compared to both dielectric lens designs 86 and 88, the passive structure generation device 20 can generate the dielectric lens shape using a polynomial with "n" order coefficients, which is essentially 2D as a representation, and rotate the polynomial 360 degrees to generate a 3D shape. The passive structure generation device 20 uses the aforementioned rotational body technique to generate both dielectric lens designs 86 and 88.

[0081] The EM simulator 30B can be run to determine a given data point for FoM data 34C. An example calculation is shown in the following formula (3):

[0082] …(3)

[0084] In some examples of the technology according to this disclosure, the EM simulator 30B can set the constant α to a value of 0.7. As can be seen from formula (3), the FoM of a phased antenna array without any dielectric lens enhancement is zero (0). Therefore, any FoM value above zero (0) represents an improvement relative to the unenhanced phased antenna array. The weighting of the gain is based on the scan angle and whether the gain of the phased antenna array enhanced with a dielectric lens is greater than the gain of the phased antenna array without any lens enhancement. If the gain of the antenna with the lens is less than the gain of the phased antenna array constructed without a lens, it will be adversely affected by a factor of five (5) regardless of the scan angle. If the gain of the phased antenna array enhanced with a dielectric lens is greater than the gain of the phased antenna array constructed without a lens at a smaller scan angle, the weighting factor is one (1). For the larger scan angle in this scenario, the weighting factor is five (5).

[0085] Additional modifications to the above calculations based on certain non-limiting examples of this disclosure include one or more of the following: (i) setting the gain to minimum_no_lens_gain for missing scan angles; (ii) setting the beamwidth to the average beamwidth (for missing values); (iii) interpolating the with_lens use case to have a resolution of +2 degrees along the scan angle axis; and / or (iv) subtracting the variance or standard deviation of the gain from the FoM value.

[0086] Therefore, the above FoM determination can be summarized according to the following five rules: (i) with lens enhancement, higher gain and lower beamwidth (especially at large scan angles) will result in a larger FoM; (ii) at any scan angle where the gain with lens enhancement becomes less than the gain without lens enhancement, FoM is severely penalized (by a weighting factor of 5); (iii) if the gain of the antenna with lens enhancement is higher than the gain of the antenna without lens enhancement at all scan angles, the gain at larger scan angles is given a greater weight than the gain at smaller scan angles; (iv) compared to the use case without lens enhancement, when the lens is placed on the antenna array, the gain enhancement with lens enhancement is given a greater weight than the reduction in beamwidth, with the former given a 70% weight and the latter a 30% weight; and (v) the smaller the change in the gain versus scan angle curve, the larger the FoM will be. These aspects are illustrated in Table 3 below:

[0087] (Option Z) Higher reward-quality factor at higher scan angles

[0088]

[0089] Table 3

[0090] Figure 10 This is a graph 90 showing one of the performance metrics 34B (i.e., peak gain) for various phased antenna arrays at various scan angles. Reference line 92 plots the peak gain of a phased antenna array without lens enhancement as the scan angle increases. Conventional lens line 94 plots the peak gain of a phased antenna array enhanced with conventional lenses as the scan angle increases. Automated design line 96 plots the peak gain of a phased antenna array enhanced with dielectric lenses designed by the passive structure design device 20 using the techniques of this disclosure as the scan angle increases. Reference line

[0091] As shown by the comparison of automated design line 96 with conventional lens line 94 and control line 92, the dielectric lens-enhanced phased antenna array designed using the Bayesian optimization automated design technique according to this disclosure provides a larger and more consistent improvement in peak gain (e.g., as expressed by beam scan range) than other scenarios. In this way, the passive structure design device 20 implements the techniques of this disclosure to generate dielectric lens designs that improve performance metric 34B (in this case, peak gain) while utilizing Bayesian optimization to reduce the computational resource expenditure of generating the dielectric lens designs. FoM data 34C reflects a FoM value of 1.03 for the dielectric lens associated with automated design line 96, while FoM data 34C reflects a FoM value of -8.49 for the conventional dielectric lens associated with conventional lens line 94.

[0092] The FoM data 34C generated by the EM simulator 30B based on one or more optimized passive structural designs 34E enhanced by the phased antenna array also shows significant improvements in storage for other types of data in the performance metric 34B. For example, in simulation runs performed by the EM simulator 30B, the phased antenna array enhanced with the aforementioned dielectric lens with a FoM value of 1.03 shows a significant reduction in beamwidth as the scan angle widens, compared to an unenhanced phased antenna array and / or a phased antenna array enhanced with a conventional dielectric lens.

[0093] Figure 11 Various aspects of the optimized passive structure design 34E compared to a conventional passive structure design are illustrated. Dielectric lens designs 98A and 100A illustrate two examples of shapes included in the optimized passive structure design 34E. Cross sections 98A and 100B show cross-sectional views of dielectric lens designs 98A and 100A, respectively. In contrast, conventional lens design 102A and cross section 102B illustrate renderings and cross-sectional views of the shapes of conventional dielectric lens designs, respectively. Figure 11 As shown, conventional lens design 102A and cross section 102B reflect hemispherical lens design.

[0094] The following is an overview of simulation results relative to designs shown, such as those using dielectric lens design 98A and cross section 98B (collectively, “Automated Passive Structure Design 98”). The dielectric constant modeled for the Automated Passive Structure Design 98 is 1.49, the loss tangent is 0.0032, the diameter is 64 mm, and the maximum height is 23 mm. These metrics were set identically for comparative experiments using lenses conforming to conventional lens designs 102A and cross section 102B (collectively, “Hemispherical Lens 102”). In these simulations, the EM simulator 30B was run iteratively with a widened scan angle until sidelobes were observed to be larger than the main lobe when plotting the beamwidth data points for performance metric 34B. For the control experiment using a phased antenna array without any dielectric lens enhancement, the sidelobes started at 36 degrees. For the phased antenna array enhanced with hemispherical lens 102, the sidelobes started at 30 degrees. For the phased antenna array enhanced with Automated Passive Structure Design 98, the sidelobes started at 36 degrees. Therefore, the EM simulator 30B demonstrates an automated passive structural design 98 to provide a significant improvement over the hemispherical lens 102 in terms of side lobe initiation.

[0095] As stated above, both the unenhanced phased antenna array and the enhanced phased antenna array utilizing automated passive structure design 98 exhibit sidelobe initiation at equal scan angles, thus providing equal baselines relative to certain data points of the performance metric 34B. Relative to peak gain (and as... Figure 10 As shown, the automated passive structure design 98 provides improvements over unenhanced phased array antennas. Furthermore, in simulations run by the EM simulator 30B, the automated passive structure design 98 is shown to provide a significant beamwidth reduction compared to the unenhanced phased array antenna at scan angles ranging from 30 to 36 degrees. Therefore, the passive structure design device 20 is experimentally demonstrated to generate optimized passive structure designs in a computationally resource-efficient manner, while providing particularly advantageous improvements for 5G antenna array enhancement and construction.

[0096] Figure 12 This is a data flow diagram (DFD) 110 illustrating an example data flow according to the technology of this disclosure. According to DFD 110, a design 104 (e.g., an initial passive structural design in an initial passive structural design 34A) is fed into an objective function 106 (which may represent a single execution pass of the EM simulator 30B). The EM simulator 30B can then populate a performance metric 34B with one or more modeled performance results of the design 104. The EM simulator 30B can also form a dataset 112 using one or more data points based on FoM data 34C assigned to the design 104.

[0097] The Bayesian optimizer 30C can feed the dataset 112 and performance data 114 into the inverse surrogate model 116. The performance data 114 can represent expected or more desired data points, which can be potentially realized using one or more final entries in the optimized passive architecture design 34E. The Bayesian optimizer 30C uses a combination of dataset 112 and performance data 114 as input. According to various aspects of this disclosure, the inverse surrogate model 116 can be associated with one or more data-driven models in machine learning, deep learning, or reinforcement learning. Example architectures that can be associated with the inverse surrogate model 116 include, but are not limited to, 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.), reinforcement learning models such as deep Q-networks (DQNs), or any other model or any combination of such models. The inverse proxy model 116 can output the selected model 118 on each iteration, and can output the best candidate 122 when convergence is achieved after running multiple iterations to optimize the passive structure design 34E.

[0098] The Bayesian optimizer 30C can be configured to optimize the characteristics of a coplanar waveguide (CPW) composed of a metallic mesh conductive material to maximize electromagnetic transmission through it for 5G frequencies while satisfying optical transparency requirements. The techniques implemented by the Bayesian optimizer 30C address several technical problems through the optimization techniques disclosed herein, such as: partially or fully concurrently utilizing different variable types; evaluating the current training state of the surrogate model; finding the global maximum while reducing or eliminating the risk of getting stuck around local maxima; and / or mitigating / eliminating the incompleteness and / or unawareness of random sampling. In experiments, the desired improvement cost function (EI(x)) shown in Equation (4) and the tradeoff parameter ξ can optionally be added. Higher values ​​of ξ will result in a higher acquisition function (in...) Figure 4 (As shown at point 58) The improved standard deviation is weighted, thereby making the optimization biased towards the exploration mode.

[0099] …(4)

[0101] And among them

[0102] Best observation;

[0103] Gaussian process mean;

[0104] Standard deviation of Gaussian process;

[0105] PDF of the standard normal distribution; and

[0106] CDF of the standard normal distribution.

[0107] In some experiments conducted according to the Bayesian optimization techniques of this disclosure, the minimum feasible product is defined as the optimization of the CPW mesh for both the signal plane and the ground plane in terms of maximizing the values ​​of S11 and S21. The parameters identified for the mesh used for optimization are the mesh offset from the plane edges, the rotation of the mesh about an axis perpendicular to the plane, the shape of the polygons tessellated to form the mesh, the length of the sides of each polygon, and the width of the lines forming the mesh. While the applications illustrated in this disclosure are specific to transparent antenna development, it should be understood that the optimization techniques described herein can also be used in the context of other types of simulation work.

[0108] With the increasing prevalence of Modeling as a Service (MaaS) and leveraging data in design, computer-automated optimization can be widely applied to all types of product development. Some experiments based on these techniques utilize dedicated libraries specifically designed to be robust and general-purpose, enabling their application to any parametric optimization problem, in this case, involving simulations for CPW development and dielectric lensing for 5G applications. The methods described in this paper are also well-suited for hyperparameter tuning in ML.

[0109] To define and combine variable types, the Bayesian optimizer 30C is configured to optimize around qualitative or "categorical" variables as well as quantitative variables. Additionally, the Bayesian optimizer 30C can restrict certain quantitative variables to discrete values ​​to account for unrelated variable ranges. By classifying parameters by type, the Bayesian optimizer 30C can leverage these classifications to provide resource-saving technical improvements such as narrowing the domain, saving simulation time, and building surrogate models to better handle different types of data. Some of the variable classifications used in experiments run according to this disclosure are described below. For the "on a continuous range" classification, if a variable is quantitative and falls within a given range, the Bayesian optimizer 30C can run process 50 based on a pre-determined understanding that process 50 should explore all possibilities within that range.

[0110] For "categorical" classification, if a variable is one of a set number of qualitative labels, the Bayesian optimizer 30C can assign such a variable as "categorical." For a variable assigned as "categorical," the Bayesian optimizer 30C can predetermine that the variable can be considered discrete and that the variable does not have an inherent relative arithmetic value. Therefore, the Bayesian optimizer 30C can eliminate any range-based processing of categorical variables and can encode variables in a way that restricts processing to values ​​(such as by using one-hot encoding). Additionally, the Bayesian optimizer 30C can use the variable as a classifier to split the domain into smaller subdomains for better evaluation.

[0111] For the "discrete" category, if the Bayesian optimizer 30C specifies a variable as discrete (e.g., as one of a predetermined set of possible values), it can iterate through the possible values ​​of that variable instead of using resources to analyze a large pool of potential values ​​within a given range. In this scenario, the Bayesian optimizer 30C restricts possible solutions based on the "discrete" specification and, as with categorical variables, can utilize a specific pool of values ​​to decompose the domain into subdomains for better evaluation. For the "combinatorial discrete" category, if the variables are discrete parameters, the Bayesian optimizer 30C can specify these variables as belonging to the "combinatorial discrete" category, for which all possible values ​​within a range should be considered as part of the optimization according to various aspects of process 50. In these scenarios, the Bayesian optimizer 30C can identify preferred values ​​for all other parameters that produce the highest value of FoM 34C, regardless of the values ​​of the combined discrete variables. In this way, the Bayesian optimizer 30C can perform robust optimization relative to parameters that are unknown or highly variable in real-world applications.

[0112] The Bayesian optimizer 30C can also implement certain aspects of this disclosure to determine whether the optimization process is stuck in a local region. In some use case scenarios, the optimization process may focus on local minima within a domain. The techniques of this disclosure enable the Bayesian optimizer 30C to provide the technical benefit of evaluating which part of the domain is of interest. According to these techniques, the Bayesian optimizer can decompose a domain into buckets or subdomains or “domain spaces” by identifying a combination of categorical and discrete variables. In some examples, the Bayesian optimizer 30C can also perform domain segmentation using quantitative variables by placing values ​​into buckets, which are then defined by different ranges or different sets of ranges. The Bayesian optimizer 30C can use two quantitative variables to form axes of planes within these smaller domain spaces. The Bayesian optimizer 30C can then evaluate the concentration of samples within each of these planes. The sample density in a given domain space is referred to herein as a “location”. The Bayesian optimizer 30C can compare the locations between domain spaces to determine relative concentrations. Within a given domain space, the Bayesian optimizer 30C can construct polygons or “convex hulls” around the samples. The Bayesian optimizer 30C can determine the location of a domain space by dividing the number of samples in each domain space by the area of ​​the convex hull around the sample. This calculation is shown by the following formula (5):

[0113] …(5)

[0115] The Bayesian optimizer 30C can use these localization values ​​to push optimization toward less explored domain spaces, such as by improving the acquisition function. For example, the Bayesian optimizer 30C can improve the acquisition function by relating the values ​​to the ξ parameter in the acquisition function. For two different optimization processes, the localization of different domain spaces changes. Both use random samples for the acquisition function, but the optimization with ξ values ​​in the acquisition function shows a more dramatic increase and earlier stabilization of the corresponding values ​​in the FoM data 34C in significantly shorter runs in the EM simulator 30B. The ξ parameter is related to the localization of the domain for each sample, and the combination of ξ parameters provides benefits associated with utilizing the acquisition function.

[0116] According to various aspects of this disclosure, the Bayesian optimizer 30C can monitor the training of the surrogate model 116. As part of the overall optimization, the surrogate model 116 undergoes continuous training and retraining. To monitor the progress of training, the Bayesian optimizer 30C evaluates the extent to which the surrogate model 116 changes between optimization iterations. This technique described herein is referred to as the "predictive discrepancy" method. According to the predictive discrepancy method, the Bayesian optimizer 30C can perform the following sequence of operations (iteration and looping where necessary):

[0117] 1. Generate a uniformly distributed sample set across domains, which is provided to the proxy model 116.

[0118] 2. Run surrogate model 116 to make predictions on the sample.

[0119] 3. In the next iteration of the optimization process, after the surrogate model 116 has been updated, the surrogate model 116 is run again to predict the same samples used in step 2 above.

[0120] 4. Evaluate the differences between the predictions made in the current iteration and those made during previous iterations.

[0121] These differences are expressed as the absolute value of the percentage difference between the current prediction and the predictions of previous iterations. The maximum, average, and range of the prediction differences are considered, as their changes between iterations can inform the Bayesian optimizer 30C of the significance of the sum of each known value on the functionality of the surrogate model 116. If the Bayesian optimizer 30C determines that no prediction difference indicator changes significantly within a certain (e.g., a threshold) number of iterations, the Bayesian optimizer 30C can determine that the surrogate model 116 is thoroughly trained or “saturated” with respect to the solution space. In the experiment, a period of significantly higher variation in predicted values ​​between iterations was observed until approximately 317 iterations. After reaching 317 iterations, the average of the prediction differences reaches a relatively stable state for the remainder of the optimization iterations. Therefore, in this particular experiment, 317 iterations represent the beginning of saturation for the surrogate model 116. The Bayesian optimizer 30C can also use the prediction differences based on each domain space to monitor the extent of the influence of a given sample on a specific aspect of the surrogate model 116 and / or the progress made in training that specific aspect of the surrogate model 116.

[0122] The Bayesian optimizer 30C can implement certain techniques of this disclosure to optimize acquisition functions (such as...). Figure 4 (As shown at point 58). In some examples, the Bayesian optimizer 30C can sample across the domain in an organized manner by utilizing particle swarm optimization (PSO). By implementing PSO, the Bayesian optimizer 30C can select random samples that can be considered as “particles” and then use a sampling function to evaluate the corresponding positions of the particles. In this way, the Bayesian optimizer 30C can determine the optimal position for each particle (“optimal individual particle position”).

[0123] Subsequently, the Bayesian optimizer 30C can formulate a vector from the current set of particle positions to the corresponding set of optimal individual particle positions. For example, the Bayesian optimizer 30C can formulate a vector from the current particle position to the corresponding single optimal particle position and use vector addition to combine the formulas for the entire set of vectors into a single vector. The Bayesian optimizer 30C can then apply the vector derived in this way to each particle to move each particle to a new position.

[0124] The Bayesian optimizer 30C can iterate this process until one or more predetermined criteria are met. This iterative process essentially reflects the sliding progress of samples along the gradient defined by the acquisition function over the domain. This process covers more domains and represents a more complete application of the acquisition function to the domain.

[0125] As an alternative to PSO, another type of sampling that forms the basis of the experimental runs of the Bayesian Optimizer 30C is called “pseudo-continuous” sampling. As used in this paper, “pseudo-continuous” sampling refers to a sampling mechanism in which step sizes are provided for all quantitative range variables. The Bayesian Optimizer 30C can then sample the domain of all values ​​of those variables at these step sizes. Between iterations of the optimization operation, the Bayesian Optimizer 30C can shift the samples by a random amount between step sizes to account for potential values ​​between the step sizes. This produces more explicit and detailed sampling precision, as well as a more uniform sampling domain. Experiments using pseudo-continuous sampling runs show that samples are typically kept in the same region of the domain in an expansive manner. Combined with PSO, the implementation of pseudo-continuous sampling for ξ-location allows the Bayesian Optimizer 30C to use pseudo-continuous sampling in a more exploratory way to find other maxima across the entire domain, while still providing the benefits of a more continuous type of sampling technique.

[0126] As described above, various exploration and exploitation techniques form the basis of the experiments run using the passive structure generation device 20. This paper discusses a comparison of exploration and exploitation techniques. A Bayesian optimizer 30C is run using the same initial sample and range set but employing these different underlying methods to perform optimization. For three different optimizations, the FoM data 34C corresponding to the samples evaluated by the objective function in each iteration, as well as the progress towards the maximum FoM, are evaluated. One is a baseline optimization (using neither PSO nor ξ), another uses PSO, and the third uses a method that correlates the ξ value of the acquisition function with the location of the domain for each sample. The distinguishing features are that the ξ-based optimization seeks to improve the acquisition function, the PSO-utilizing optimization seeks to improve the acquisition function, and the baseline serves as a control experiment.

[0127] The baseline (control experiment) produced local maxima comparable to other baselines at 365 iterations. The optimization with localization inserted into ξ produced a similar global maximum only in 151 iterations, while finding many local maxima along the way. The optimization using PSO achieved a similarly high maximum FoM only in fifteen (15) iterations, but then reached a steady state and did not reach a peak FoM thereafter. Although the PSO technique explored the domain more extensively than the other two techniques, it did not enable the Bayesian optimizer 30C to find local maxima as high as those revealed by the other techniques along the way.

[0128] In the detailed description of exemplary embodiments of the invention, reference is made to the accompanying drawings, which illustrate specific embodiments in which the invention may be practiced. These exemplary embodiments are not intended to exhaustively list all embodiments according to the invention. It should be understood that other embodiments may be utilized, and structural or logical changes may be made, without departing from the scope of the invention. Therefore, the following detailed description should not be considered limiting, and the scope of the invention is defined by the appended claims.

[0129] Unless otherwise specified, all numbers used in this specification and claims to express characteristic dimensions, quantities, and physical properties should in all cases be understood to be modified by the terms “about,” “approximately,” or “substantially.” Therefore, unless stated to the contrary, the numerical parameters listed in the foregoing specification and appended claims are approximations, which may vary depending on the desired properties sought by those skilled in the art using the teachings disclosed herein.

[0130] Unless otherwise expressly stated in the description and appended claims, the singular forms “a” and “described” as used in the specification and appended claims cover embodiments having multiple referents. Unless otherwise expressly stated in the description and appended claims, the term “or” is generally used in its meaning including “and / or”.

[0131] It should be recognized that, based on this example, certain actions or events of any of the methods described herein may be performed in a different order, and may be added together, combined, or omitted (e.g., not all described actions or events are necessary for the practice of the method). Furthermore, in some examples, actions or events may be executed simultaneously rather than sequentially, for example, through multithreaded processing, interrupt handling, or multiple processors.

[0132] The techniques described in this disclosure can be implemented, at least in part, in hardware, software, firmware, or any combination thereof. For example, aspects of the described techniques can be implemented in one or more processors, including one or more microprocessors, CPUs, GPUs, DSPs, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or any other equivalent integrated or discrete logic circuits, and any combination of such components. The terms "processor" or "processing circuitry" can generally refer to any of the aforementioned logic circuits, alone or in combination with other logic circuits, or any other equivalent circuitry. A control unit containing hardware can also perform one or more of the techniques disclosed herein.

[0133] Such hardware, software, and firmware may be implemented within the same device or in different devices to support the various operations and functions described in this disclosure. Furthermore, any described unit, module, or component may be implemented together or individually as discrete but cooperating logical devices. Describing different features as modules or units is intended to highlight different functional aspects and does not necessarily imply that such modules or units must be implemented by separate hardware or software components. Rather, the functionality associated with one or more modules or units may be performed by separate hardware or software components or integrated into common or separate hardware or software components.

[0134] The techniques described in this disclosure may also be embodied or encoded in a computer-readable medium, such as a computer-readable storage medium, containing instructions. For example, when the instructions are executed, instructions embedded or encoded in a computer-readable storage medium may cause 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 read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, hard disk, CD-ROM, floppy disk, magnetic tape cassette, magnetic media, optical media, or other computer-readable media.

[0135] Various examples have been described. These examples, as well as others, are all within the scope of the following claims.

Claims

1. A passive structure generation device, the passive structure generation device comprising: At least one non-transitory computer-readable storage medium, on which instructions are stored; At least one processor, coupled to the at least one non-transitory computer-readable storage medium and configured to execute the instructions to: Receiving passive structures; The quality factor is generated using simulation techniques based on the passive structure and frequency data. A new passive structure is calculated using a Bayesian method based on the passive structure and the quality factor; and The new passive structure is output to at least one of a user interface, an external device, or at least one of the at least one non-transitory computer-readable storage media.

2. The passive structure generation device according to claim 1, wherein calculating the new passive structure further includes: A second passive structure is selected from a predetermined set of passive structures, wherein the passive structure is included in the predetermined set of passive structures.

3. The passive structure generation device according to claim 1, wherein the passive structure is previously generated based on a plurality of passive structures.

4. The passive structure generation apparatus of claim 1, wherein the passive structure comprises a three-dimensional shape generated by rotating a 2D shape associated with the coefficients of an nth-order polynomial.

5. The passive structure generation device according to claim 4, wherein the nth order is at least three.

6. The passive structure generation apparatus of claim 1, wherein the passive structure is previously generated according to an autoencoder generation method or at least one of a set of parameters.

7. The passive structure generation device according to claim 1, wherein calculating the new passive structure comprises: Generate a sample pool; Use the acquisition function to select the final sample from the sample pool; and The final sample is provided to the simulation technique to generate a quality factor, and the final sample is associated with the new passive structure.

8. The passive structure generation device of claim 7, wherein the sample comprises a set of parameter values ​​associated with the new passive structure.

9. The passive structure generation device according to claim 8, wherein the set of parameter values ​​includes at least one of width value, spacing value, shape value and rotation angle value.

10. The passive structure generation device according to claim 8, wherein the set of parameter values ​​includes an abstract feature representation.

11. The passive structure generation device according to claim 7, wherein the quality factor includes multiple performance parameter values.

12. The passive structure generation device according to claim 7, wherein calculating the new passive structure further includes: The surrogate model is updated based on the quality factor and the final sample.

13. The passive structure generation device of claim 12, wherein the proxy model comprises a function associated with at least two parameters.

14. The passive structure generation device according to claim 13, wherein the at least two parameters include spacing and width.

15. The passive structure generation device according to claim 13, wherein the function includes at least one conditional statement.

16. The passive structure generation device of claim 15, wherein the at least one conditional statement includes a minimum transparency value, a maximum lens diameter, or a maximum lens height.

17. The passive structure generation device according to claim 15, wherein the at least one conditional statement comprises a relationship between two or more boundary parameters.

18. The passive structure generation device according to claim 7, wherein generating the sample pool includes randomly generating the sample pool.

19. The passive structure generation device according to claim 7, wherein generating the sample pool comprises: The sample pool is generated based on a proxy function.

20. The passive structure generation device according to claim 7, wherein the sample comprises a set of parameter values, and generating the sample pool comprises: Stepping is performed on at least one parameter of the set of parameter values ​​of samples included in a previously generated sample pool in at least one direction.

Images