Passive structural design for phased antenna array

Abstract

A method of generating a passive fabric design includes receiving, by processing circuitry of a computing device, a set of performance metrics. The method also includes providing, by the processing circuitry, the set of performance metrics to a trained neural network. The method also includes receiving, by the processing circuitry, a passive structure design associated with a passive structure from the trained neural network. The method also includes outputting, by the processing circuitry, the passive fabric design via a communication interface.

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 that comply with the fifth generation technical standards for broadband cellular networks. Each such site typically supports three sector antenna arrays, where each antenna array provides 120° (i.e., + / - 60°) azimuth coverage within a given cell. In combination, the three sector antennas provide 360° omnidirectional coverage around the site. Conventional 4G / LTE systems operate at frequencies below 6 GHz, which have low propagation losses compared to mmWave frequencies that are typically at or above 20 GHz. To provide network coverage at the same range (distance from the RAN cell site) as conventional 4G / LTE systems, mmWave antennas must (i) be highly directional and (ii) have a steerable radiation pattern. Antenna arrays including multiple radiating elements provide these enhancements in the context of mmWave equipment. Antenna arrays typically use a massive multiple input multiple output (massive-MIMO) architecture to provide spatial diversity (an aspect by which a base station can communicate with multiple devices within the same cell using the same time-frequency resources with the help of highly directional antennas), as in the case of 5G system specifications.

[0003] However, high directivity introduces one or more trade-offs. Providing high directivity requires many elements in each phased array and limits the azimuth scan range of the entire antenna system due to beam broadening as the phased array broadcasts further away from the optical axis (referred to as the "antenna boresight"). Thus, highly directional mmWave phased arrays cannot provide 120° coverage at wide scan angles without introducing significant gain degradation. This gain degradation results in reduced coverage at the sector seams within a communication cell.

[0004] The techniques of this disclosure relate to designing passive structures (e.g., dielectric lenses) that widen the scan range of mmWave phased antenna arrays. Potential advantages provided by the passive structure designs (e.g., dielectric lens designs) of this disclosure relate to obtaining a narrower beamwidth. For example, a lower order array (e.g., an array with a smaller number of antenna elements) incorporating the passive structure designs of this disclosure can provide a resolution similar to that of a higher order array (e.g., an array with a larger number of antenna elements).

[0005] In one example, a system includes: interface hardware; a memory communicatively coupled to the interface hardware; and processing circuitry communicatively coupled to the memory and the interface hardware. The memory is configured to store a set of performance metrics. The processing circuitry is configured to provide the set of performance metrics to a trained neural network, and to receive from the trained neural network a passive structure design associated with a passive structure. The interface hardware is configured to output the passive structure design.

[0006] In another example, a method includes receiving, by processing circuitry of a computing device, a set of performance metrics, and providing, by the processing circuitry, the set of performance metrics to a trained neural network. The method further includes receiving, by the processing circuitry, from the trained neural network a passive structure design associated with a passive structure, and outputting, by the processing circuitry, the passive structure design via a communication interface.

[0007] In another example, a device includes: means for receiving a set of performance metrics; means for providing the set of performance metrics to a trained neural network; means for receiving from the trained neural network a passive structure design associated with a passive structure; and means for outputting the passive structure design.

[0008] In another example, a device includes means for training a first neural network to generate a latent space representation based on a passive structure design, the first neural network including a first encoder and a first decoder. The device further includes means for training a second neural network to generate an output passive structure design based on performance metrics, the second neural network including a second encoder and a second decoder, the second neural network including the first decoder.

[0009] In another example, a method of training a neural network includes training a first neural network to generate a latent space representation based on a passive structure design, the first neural network including a first encoder and a first decoder. The method further includes training a second neural network to generate an output passive structure design based on performance metrics, the second neural network including a second encoder and a second decoder, the second neural network including the first decoder.

[0010] In another example, a system includes a memory and processing circuitry communicatively coupled to the memory. The memory is configured to store performance metrics. The processing circuitry is configured to train a first neural network to generate a latent space representation based on a passive structure design, the first neural network including a first encoder and a first decoder. The processing circuitry is further configured to train a second neural network to generate an output passive structure design based on the performance metrics stored to the memory, the second neural network including a second encoder and a second decoder, the second neural network including the first decoder.

[0011] In another example, a non-transitory computer-readable storage medium is encoded with instructions. When executed by one or more processors, the instructions cause the one or more processors to: receive a set of performance metrics; provide the set of performance metrics to a trained neural network; receive from the trained neural network a passive structure design associated with a passive structure; and output the passive structure design via a communication interface.

[0012] The passive structure design techniques of the present disclosure provide several technical improvements in the technical field of phased antenna array design. In this way, the passive structure design of the present disclosure can reduce the design complexity and cost of a phased antenna array system 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 the present disclosure provide these performance enhancements in the context of a three-sector antenna implementation, but in many cases, the infrastructure can be reduced to a single antenna or a two-antenna array, particularly in use cases covering smaller, more densely populated device deployment areas such as urban downtown areas. Additionally, the improvements provided by the passive structures designed according to the techniques of the present disclosure improve cell-level performance and capacity, potentially reducing the number of arrays required from a higher (e.g., system-level) perspective. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1A and Figure 1B illustrate the difference between the signal coverage provided by an existing phased antenna array and the enhanced signal coverage of a corresponding phased antenna array equipped with a passive structure designed according to the techniques of the present disclosure.

[0014] Figure 2 is a flowchart illustrating an example workflow of the present disclosure.

[0015] Figure 3 is a block diagram illustrating an example implementation of a passive structure generation device of the present disclosure.

[0016] Figure 4A and Figure 4B respectively illustrate a three-dimensional representation and a two-dimensional representation of a dielectric lens generated by a generative adversarial network (GAN)-based implementation of a trained model of the present disclosure when run during an execution phase.

[0017] Figure 5 is illustrative of a phased antenna array incorporating dielectric lenses generated by combining the GAN designs shown in Figure 4A and Figure 4B showing a graph of the gain values provided.

[0018] Figure 6A and Figure 6BIllustrate three-dimensional representations and two-dimensional representations of dielectric lenses generated by a GAN-based implementation of a trained model of the present disclosure when run during an execution phase, respectively.

[0019] Figure 7 Is a graph illustrating the gain values provided by a phased antenna array incorporating a dielectric lens designed by combining the GAN generations shown in Figure 6A And Figure 6B

[0020] Figure 8 Is a graph showing the increase in peak gain (or maximum gain) provided by a phased antenna array equipped with a dielectric lens of the inverse design of the present disclosure compared to the peak gain provided by other phased antenna arrays.

[0021] Figure 9 Is a graph comparing the increase in peak gain (or maximum gain) provided by phased antenna arrays equipped with modified and unmodified versions of the dielectric lens of the inverse design of the present disclosure and by phased antenna arrays not enhanced by any dielectric lens infrastructure.

[0022] Figure 10 Is a block diagram illustrating aspects of the training phase of a training model with respect to Figure 3

[0023] Figure 11 Is a block diagram illustrating an example of the architecture of a training model and a trained model based on an autoencoder neural network (ANN). Figure 3

[0024] Figure 12 Illustrates an example of a dielectric lens design generated by an ANN-based implementation of a trained model of Figure 3

[0025] Figure 13 Illustrates an example of a dielectric lens design generated by an ANN-based implementation of a trained model of Figure 3 From a top view.

[0026] Figure 14A Is a scatter plot clustering fully encoded instances of the latent space representation of various lens designs. Figure 14B Is a conceptual diagram of a top view illustrating the selection of a lens design represented in Figure 14A

[0027] Figure 15 Illustrates the results of the GAN-based inverse design optimization technique of the present disclosure.

[0028] Figure 16 Illustrates an example of an interpolated lens design generated as part of the labeled dataset generation technique of the present disclosure.​​​​​ Detailed implementation manners

[0029] The system of the present disclosure solves various performance defects of existing 4G / LTE antenna arrays when reused for 5G voice and data service delivery. A passive structure (e.g., a dielectric lens) designed according to the technology of the present disclosure improves signal coverage when incorporated into a phased antenna array. For example, a passive structure designed according to the technology described herein can enable a phased antenna array to maintain a non-increasing beam width even when the azimuth angle increases.

[0030] Figure 1A and Figure 1B illustrates the difference between the signal coverage provided by an existing phased antenna array and the enhanced signal coverage of a corresponding phased antenna array equipped with a passive structure designed according to the technology of the present disclosure. Figure 1A illustrates the signal coverage provided by a three-sector antenna array equipped with a currently available passive structure (e.g., a dielectric lens). Figure 1A The system 10A provides three instances of radio frequency (RF) beams at boresight angles (i.e., zero degrees with respect to the respective antennas of the three-sector array) that are equally spaced 120 degrees from each other and reach the cellular boundary 12. Figure 1A illustrates a signal blind spot 14, which is a non-limiting example of a signal coverage gap (or so-called "dead zone"). In addition to the signal blind spot 14, the system 10A includes two other signal coverage gaps. The RF beams in these regions of the signal coverage gap exhibit a widened beam width proportional to the corresponding azimuth angle. The wider beam width associated with these RF beams reduces the corresponding beam depth, thereby resulting in a reduction in signal coverage range or a dead zone before reaching the perimeter represented by the cellular boundary 12.

[0031] Figure 1B illustrates the 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 the present disclosure. Figure 1B All sixteen RF beams of the system 10B reach the cellular boundary 12 because the beam width of the beams does not increase with the azimuth angle and the gain is uniform with respect to the azimuth angle. The radiation pattern of the boresight provides signal coverage throughout the 360-degree scan of the system 10B, as shown by the non-limiting example of the signal coverage area 16.

[0032] In this way, a passive structure (e.g., a dielectric lens) designed according to the techniques of the present disclosure achieves 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 by a passive structure that is designed by the reverse design techniques of the present disclosure. The reverse design techniques of the present disclosure take one or more performance metrics as input. The performance metrics can represent any one of the minimum performance criteria of a planned phased antenna array, the maximum possible performance of a planned phased antenna array, its average value, or any other planned performance criteria. In one non-limiting example, Figure 1B one or more aspects of the signal coverage shown in System 10B of Figure 1B can be represented in the performance metrics.

[0033] The performance metrics can include training-phase inputs, execution-phase inputs, or both, or be them, or be part of them. In various examples consistent with the present disclosure, the performance metrics can be included in the input data provided to a machine learning (ML) model or an artificial intelligence (AI) model. During the training phase of the AI / ML model, the systems of the present disclosure can train the AI / ML model by providing a set of performance metrics as training data. During the execution phase of the trained AI / ML model, the device can provide the set of performance metrics as execution-phase inputs to the trained AI / ML model. Whether during the training phase or the execution phase, the AI / ML model can return a passive structure design as output. In various non-limiting examples, the AI / ML model can return a passive structure design representing the planned structure of a dielectric lens, the planned structure of an irregular-shaped lens, etc.

[0034] According to the automated design techniques of the present disclosure, various AI / ML model architectures can be used. For example, various types of neural networks are compatible with the automated design techniques of the present disclosure. In some examples, the trained AI / ML model can represent a generator network that is trained as part of a generative adversarial network (GAN) that uses a complementary discriminator network during the training phase. In various examples of a GAN-based implementation of the techniques of the present disclosure, the generator network and the discriminator network are included in at least one of a deep convolutional generative adversarial network (DCGAN), a Wasserstein generative adversarial network (WassersteinGAN), a PixelGAN, or a CycleGAN. In some examples, the trained generator can be an autoencoder model. An example of an autoencoder model that can be used according to the reverse design techniques of the present disclosure is a variational autoencoder (VAE).

[0035] Figure 2is a flowchart illustrating an example workflow 20 of the present disclosure. Workflow 20 includes a training phase 14 and an execution phase 16. It will be understood that workflow 20 may be implemented locally on a single system or in a distributed manner across different systems. In a distributed implementation, the training phase 14 may be implemented by a first computing system, and a trained model that is an output of the training phase 14 may be placed in the execution phase 16 by a second computing system different from the first computing system. Although, for ease of illustration, the execution phase 16 is shown by a single instance in Figure 2 it should be understood that the execution phase 16 may be implemented in multiple instances. For example, multiple instances of the trained model output by the training phase 14 may be deployed across multiple computing systems, thereby enabling the multiple computing systems to individually generate passive structure designs as needed while leveraging the inverse design techniques of the present disclosure in each instance.

[0036] The training phase 14 of workflow 20 may begin by collecting one or more performance metrics (18). For example, the training system of the present disclosure may collect performance metrics of lenses of various shapes, form factors, etc., where the performance metrics indicate aspects of a phased antenna array equipped with a lens of each type (shape and / or form factor). The training system may maintain a corresponding one-to-one mapping between the respective performance data (e.g., a single performance metric or a discrete grouping of performance metrics) and the type of lens from which the respective performance data was obtained (the latter is also referred to herein as a “training passive structure”).

[0037] The training system may use training data pairs to train an AI / ML model (22). In an example where the model conforms to the GAN architecture, the training system may provide the corresponding performance data to the generator network of the GAN on a per-training-pair basis and receive an artificial passive structure as an output from the generator network. Subsequently, the training system may provide a combination of the passive structure (which is included in the training pair) and the artificial passive structure (which is output by the generator network) to the discriminator network of the complementary generator network in the GAN architecture. The training system may receive a score from the discriminator network. Using the score returned by the discriminator network, the training system may update the generator network and the discriminator network. The training system may iterate the above training operations for the generator network and the discriminator network of the GAN until the training system determines that the generator network has achieved convergence.

[0038] The training phase 14 of workflow 20 can end with the trained system outputting a trained AI / ML model (24). In the GAN-based example above, the training system can discard the trained discriminator network and only output the trained generator network. As described above, the training system can output one or more instances of the trained generator network based on need. In this way, the training phase 14 of workflow 20 supports scalability by providing the ability to deploy the trained generator network to a potentially large number of systems (hereinafter, "utilization systems") that can execute the trained generator network by providing desired performance metrics as input to generate passive structure designs to be used in a phased antenna array configuration. In some use case scenarios, the training system can also be one of the utilization systems that execute the trained model.

[0039] The execution phase 16 of workflow 20 can be implemented by any one or more of the utilization systems to which the trained model is deployed, as described in the conclusion above with respect to the training phase 14. The execution phase 16 can begin with the respective utilization system providing one or more expected performance metrics (26) to the trained model. The expected performance metrics can represent one or more of a gain value (e.g., maximum gain typically corresponding to beam depth), beam width (e.g., half-power beam width), etc. In some examples, each expected performance metric provided as an input to the execution phase is associated with a corresponding scan angle. For example, each scan angle can represent an increment in azimuth angle using the line-of-sight axis as a baseline.

[0040] The utilization system can obtain a passive structure design (28) as an output from the trained model. In various examples consistent with the present disclosure, the passive structure design returned by the trained model at the end of the execution phase 16 can represent the shape of a dielectric lens to be used in a phased antenna array. Since the trained model of the present disclosure takes performance metrics as inputs in the execution phase 16 and outputs a dielectric lens design, the techniques implemented by the trained model of the present disclosure are referred to as "reverse design" with respect to the dielectric lens or other passive structures.

[0041] The results of experiments using the passive structure designs of the present disclosure are described below, as well as the technical improvements observed from the experimental results. A dielectric lens manufactured according to the passive structure design output by the trained model of the present disclosure is assembled into a 4×4 antenna array to verify the design. A 4×4 antenna array typically provides 60 degrees (i.e., + / -30 degrees) azimuth coverage, while an 8×8 antenna array typically provides 120 degrees (i.e., + / -60 degrees) azimuth coverage. For various reasons, a 4×4 antenna array is used instead of an 8×8 antenna array to verify the dielectric lens manufactured according to the reverse design specifications generated by the trained model of the present disclosure, some of which reasons are described below.

[0042] As an example, compared to an 8×8 antenna array, it is easier to control individual antenna elements (e.g., amplitude and / or phase) in a 4×4 antenna array. The incorporation of an external lens structure makes it more important to have control over individual antenna elements to control the scan angle. As another example, the execution phase 16 runs faster (sometimes seven times faster) with respect to the passive structure design for a 4×4 array than for an 8×8 array, and the dielectric lens design generated for a 4×4 antenna array can be extrapolated to an 8×8 antenna array relatively easily. Since the goal of the experiment is to validate the lens concept via models and measurements, this type of extrapolation is feasible. Contrary to the 8×8 antenna array scenario, for the 4×4 antenna array scenario, the number of iterations required for the training phase 14 as well as the execution phase 16 is significantly reduced. In some scenarios, the trained model can generate the inverse design of the dielectric lens within seconds.

[0043] Figure 3 is a block diagram illustrating an example specific implementation of the passive structure generation device 30 of the present disclosure. Although Figure 3 illustrates one specific implementation of the passive structure generation device 30 consistent with aspects of the present disclosure, it will be understood that other architectures (whether single-device architectures or distributed architectures) of the functionality described with respect to the passive structure generation device 30 are also consistent with aspects of the present disclosure. In Figure 3 a particular example of, the passive structure generation device 30 performs the functionality of both a training system (which performs the Figure 2 training phase 14) and a utilization system (which performs the Figure 2 execution phase 16). In other examples consistent with the present disclosure, the system can be configured to perform specifically as a training system or specifically as a utilization system. Thus, it will be appreciated that different configurations are consistent with the techniques of the present disclosure, and Figure 3 illustrates a non-limiting example of a system configuration consistent with the present disclosure.

[0044] In Figure 3In the example, the passive structure generation device 30 includes a processing circuit 31 and a memory 32. In some examples, the processing circuit 31 and the memory 32 may be integrated into a single hardware unit such as a system-on-chip (SoC) or an integrated circuit (IC). The processing circuit 31 may represent one or more processors or processing units, each of which may include a multi-core processor, a controller, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a processing circuit (e.g., fixed-function circuitry, programmable circuitry, or any combination of fixed-function circuitry and programmable circuitry), or one or more of equivalent discrete logic circuitry or integrated logic circuitry. The memory 32 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 embodiments, the memory 32 may include a single memory unit or multiple memory units.

[0045] The combination of the memory 32 and the processing circuit 31 provides a computing platform for executing the operating system 24. The operating system 24 provides a multitasking operating environment for executing one or more software components 42. As shown, the processing circuit 31 is connected to external systems and devices via an input / output (I / O) interface 33, e.g., via one or more communication networks. The I / O interface 33 may incorporate network interface hardware, such as one or more wired and / or wireless network interface controllers (NICs) for communicating via a communication link 38.

[0046] In Figure 3 a particular example, the communication link 38 represents one or more network-enabled communication connections, such as links to one or more packet-switched networks collectively referred to as "communication networks" in the Figure 3 context. The communication network may represent 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.

[0047] The communication link 38 communicatively couples the passive structure generation device 30 to the communication network and, via the communication network, to other devices. Each of the communication links 38 may include one or more wired connections (e.g., connections), wireless connections (e.g., Wi-Fi TM connections), or a combination of both wired and wireless communication connections. In Figure 3In the illustrated example, I / O interface 33 also facilitates communication between the passive structure generation device 30 and one or more remote devices 34 via a communication link 38. In Figure 3 a specific example, the communication link 38 represents one or more local connections, such as connections to the remote device 34 via a local area network (LAN) and / or a personal area network (PAN), such as near field communication (NFC) pairing, pairing, pairing, etc. The remote device 34 may include any one or more of a computing device (e.g., a device deployed at a dielectric lens manufacturing entity), an additive manufacturing device (e.g., a so-called "3D printer"), etc.

[0048] In Figure 3 the specific implementation shown, the bus 40 provides inter-component connectivity between the processing circuit 31, the memory 32, and the I / O interface 33. The bus 40 may represent a half-duplex or full-duplex bus that provides data transfer capabilities between two or more of the processing circuit 31, the memory 32, the I / O interface 33, and / or any other hardware components of the passive structure generation device 30. The bus 40 may represent various types of system buses or computer buses, including one or more bus networks. Regardless of the implemented topology, in various examples, the bus 40 may incorporate various types of inter-component connectivity hardware, such as those that conform to any one of the first-generation, second-generation, third-generation, or fourth-generation bus or bus network technologies as set forth by the Institute of Electrical and Electronics Engineers (IEEE) and / or other bus or bus network technologies defined in standards that are being developed or will be adopted later.

[0049] In Figure 3 a specific example, the software components 42 of the passive structure generation device 30 include a training unit 42A, a design unit 42B, and a deployment unit 42C. In some exemplary methods, one or more of the software components 42 represent executable software instructions that may take the form of one or more software applications, software packages, software libraries, hardware drivers, and / or application programming interfaces (APIs). Additionally, any of the software components 42, when executed, may cause the passive structure generation device 30 to output data and / or receive data via the I / O interface 33.

[0050] The aspect of the memory 32 that provides non-volatile storage and / or long-term storage supports the local storage of the data repository 44 at the passive structure generation device 30. In Figure 3In the example of, the data repository 44 includes performance metrics 44A, trained passive structure designs 44B, training data pairs 44C, expected performance metrics 44D, and output passive structure designs 44E. One or more of the software components 42 may call the processing circuitry 31 and the memory 32 to access one or more of the data repositories in the data repository 44 to retrieve data for various purposes, such as for model training, trained model deployment, trained model execution, or passive structure design deployment.

[0051] In some examples, the software component 42 may implement read / write capabilities with respect to the data repository 44, such as accessing and using information obtainable from the data repository 44 and / or modifying information currently stored in the data repository 44. In embodiments where the passive structure generation device 30 represents a distributed computing system, one or more of the data repositories in the data repository 44 may be located partially or entirely at a remote location from the processing circuitry 31, and in such embodiments, the software component 42 may use the NIC hardware of the I / O interface 33 to access the data repository 44.

[0052] The passive structure generation device 30 may obtain the performance metrics 44A and / or the trained passive structure designs 44B from one or more sources (such as from the remote device 34) through direct user input entered using one or more input devices coupled to the passive structure generation device 30 via the I / O interface 33. The trained passive structure designs 44B may include one or both of a design representing a dielectric lens that has been used in the field or in an experimental environment and an expected design representing a dielectric lens that has not yet been manufactured or prototyped. The performance metrics 44A may include performance indicator data collected from experiments conducted using dielectric lenses in one or more configurations.

[0053] In some examples, the performance metrics 44A may include simulation-generated performance indicators for a dielectric lens design for which performance is modeled via a simulator or simulation technique. Non-limiting examples of the performance metrics 44A include values such as gain, beam width (in some instances, which may be specified as "half-power beam width"), maximum gain (which is an indicator of beam depth), etc. In various use case scenarios consistent with aspects of the present disclosure, each data point included in the performance metrics 44A may be matched with a specific scan angle at which the corresponding performance metric was observed or simulated.

[0054] In various experiments, the performance metric 44A is populated with finite-difference time-domain simulations performed using a 3D electromagnetic (EM) analysis software suite, where each corresponding simulation is characterized by a different dielectric lens placed on a 4×4 element chip array that is modeled after a commercially available 28 GHz array. The candidate dielectric lenses are subject to specific size constraints and are designated to have a relative permittivity (∈r) value of 1.5 and a loss tangent of 0.0032. In some instances, in addition to the fully simulated data points obtained in this manner, the passive structure generation device 30 may use data points generated by an autoencoder to enhance the performance metric 44A.

[0055] For example, the passive structure generation device 30 may execute an autoencoder that is trained via a learning mechanism to encode dielectric lens designs into a latent space embedding. The passive structure generation device 30 may then interpolate the latent space embedding to generate various dielectric lens shapes. Empirical observations indicate that the combined performance of the offspring lenses (i.e., dielectric lens designs obtained from interpolating two given parent lenses) is relatively close to the interpolated performance of the parent lenses.

[0056] In some examples, the passive structure generation device 30 may utilize various techniques enhanced by a variational autoencoder. The passive structure generation device 30 may use the variational autoencoder to generate dielectric lens shapes for training the passive structure design 44B. For example, the passive structure generation device 30 may utilize a combination of data accuracy and dimensionality reduction capabilities provided by the variational autoencoder. By leveraging the scalability provided by the latent space dimensionality reduction on the encoder side of the variational autoencoder and the low-loss reconstruction provided by the decoder side of the variational autoencoder, a dielectric lens design with accurate performance is delivered to the training passive structure design 44B in a computationally lightweight manner.

[0057] The passive structure generation device 30 may invoke the training unit 42A to form training data pairs 44C. For example, the training unit 42A may match each corresponding scan angle-specific metric of the performance metric 44A with a corresponding pre-determined dielectric lens design of the training passive structure design 44B to form a corresponding pair of the training data pair 44C. In a simulation-driven example, the training unit 42A may pair dielectric lens designs generated by an autoencoder (e.g., a variational autoencoder) with performance metrics generated by 3D EM analysis software to form a corresponding pair of the training data pair 44C. Each pair of the training data pair 44C can generally be represented in the following format: {engineered design, performance}.

[0058] In Figure 3In the illustrated specific implementation, the passive structure generation device 30 includes an AI / ML model 46. The AI / ML model 46 includes a training model 46A and a trained model 46B. The training unit 42A can use the training data pair 44C as training data to train the training model 46A. That is, the training unit 42A can train the training model 46A during the training phase 14 to provide various pairs of the training data pair 44B to the training model 46A as training data. Each pair of the training data pair 44B can include a single entry (representing a dielectric lens design) from the training passive structure design 44B and corresponding performance data, which can represent one or more data points in the performance metric 44A. Each of the training data pairs 44C provided by the training unit 42A to the training model 46A is associated with a corresponding scan angle at which the performance data reflects the performance delivered by the corresponding dielectric lens when integrated into the phased antenna array. In some examples, the passive structure generation device 30 (and / or other systems configured to perform the techniques of the present disclosure) can use Bayesian techniques to generate each tuple of the training passive structure design 44B.

[0059] In some specific implementations according to the techniques of the present disclosure, the training model 46A represents a neural network, such as an autoencoder neural network (ANN). The techniques of the present disclosure incorporate optimization into the ANN via the derivation of the latent space representation while also leveraging the interpolation and extrapolation design capabilities provided by the ANN. The techniques of the present disclosure utilize an optimization method with a target reward function for directed design experiments as part of the inverse generative design functionality described herein. In an instance where the training model 46A represents an ANN, the training model 46A includes two neural networks, namely an encoder network and a decoder network. The encoder network and the decoder network of the ANN utilize dimensionality reduction capabilities that enable the two networks to operate on a reduced-dimensional vector representation referred to as a "latent space representation". Although the techniques of the present disclosure are described with respect to the use case of EM device design, it should be understood that the optimization-enhanced ANN application techniques of the present disclosure can also be applied to other end-use applications.

[0060] In some specific implementations according to the techniques of the present disclosure, the training model 46A represents an artificial neural network, such as a variational autoencoder (VAE) network. In the VAE-based specific implementation, the techniques of the present disclosure utilize significant encoder-side dimensionality reduction and low-loss (or potentially even lossless) decoder-side reconstruction in the formation of the latent space representation. In this way, in the VAE-based specific implementation, the passive structure generation device 30 reduces the computational resource occupancy of the inverse design techniques of the present disclosure while improving data accuracy through the low-loss or potentially lossless reconstruction provided by the VAE.

[0061] In some specific implementations of the technology according to the present disclosure, the trained model 46A represents an adversarial learning-based model using two or more neural networks, such as a generative adversarial network (GAN). In an instance where the trained model 46A represents a GAN, the trained model 46A may include two neural networks, namely, a generator network and a discriminator network. In a specific implementation based on a GAN, the trained model 46A may represent any one of various types of GANs, such as a deep convolutional generative adversarial network (DCGAN), a Wasserstein generative adversarial network (Wasserstein GAN), a PixelGAN, a CycleGAN, or any other type of GAN. In some GAN-based specific implementations of the trained model 46A, the generator network may represent a U-Net. In some GAN-based specific implementations of the trained model 46A, the generator network and / or the discriminator network may represent corresponding autoencoder networks.

[0062] The training unit 42A may perform training and retraining operations on the trained model 46A using the training data pair 44B until the trained model 46A achieves convergence. When it is detected that the trained model 46A has achieved convergence, the training unit 42A may output the trained model 46B and store the trained model 46B in the AI / ML model 46. In the case of a GAN-based specific implementation, the trained model 46A may output the trained generator network of the GAN as the trained model 46B. In the case of ANN-based and VAE-based specific implementations, the trained model 46B may represent a combination of an encoder network and a decoder network with trained generated encoding weights and trained generated decoding weights.

[0063] The design unit 42B may place the trained model 46B in the execution phase 16 to run the trained model 46B. In various use case scenarios of the present disclosure, the design unit 42B may provide one or more data points of the expected performance metric 44D as an execution phase input to the trained model 46B. That is, the design unit 42B may implement aspects of the reverse design-based present disclosure by providing one or more of the expected performance metrics 44D (such as performance goals, minimum performance requirements, or other types of expected performance indications) as inputs to the trained model 46B during the execution phase 16. Subsequently, the trained model 46B may terminate at least one pass of the execution phase 16 by outputting a passive structure design. In this way, the design unit 42B implements the reverse design technique of the present disclosure to run the trained model 46B by providing a portion of the expected performance metric 44D as an input and obtaining a passive structure design, and the trained model 46B uses the input portion of the expected performance metric 44D as a performance guide for the corresponding passive structure to generate the passive structure design.

[0064] In some examples, the design unit 42B may store the passive structure design obtained from the execution phase 16 of the trained model 46B to output the passive structure design 44E. In some examples, the deployment unit 42C may call the I / O interface 33 to output the passive structure design obtained from the execution phase 16 of the trained model 46B to an external device, such as outputting to one or more of the remote devices 34 via the communication link 38. In some examples, the design unit 42B may save the passive structure design obtained from the trained model 46B to output the passive structure design 44E, and the deployment unit 42C may use the I / O interface 33 to output the passive structure design obtained from the trained model 46B to an external device.

[0065] In some examples, the deployment unit 42C may call the I / O interface 33 to signal data representing one or more pre-saved passive structure designs that output the passive structure design 44E. In these examples, the passive structure generation device 30 conveys to one or more external devices the passive structure designs previously generated by the trained model 46A using certain parameters of the expected performance metric 44D.

[0066] For example, the deployment unit 42C may select the design of a particular output passive structure design 44E that is generated using those expected performance metrics of the expected performance metric 44D that match or are equivalent to certain expected performance parameters received from a user of the passive structure generation device 30 via the I / O interface 33 or from an external device communicatively coupled to the passive structure generation device 30 via the communication link 38. In this way, the deployment unit 42C may implement the techniques of the present disclosure such that the passive structure generation device 30 can provide previously generated passive structure designs in response to performance parameter-based requests received from the user and / or external device.

[0067] In some examples, the deployment unit 42C may call the I / O interface 33 to deploy the trained model 46B to an external device, such as signaling the trained model 46B to the remote device 34 via the communication link 38. In some examples, the deployment unit 42C may call the I / O interface 33 to save a copy of the trained model 46B to a removable storage device 52. The removable storage device 52 may represent any type of non-volatile storage medium, such as an external hard drive or solid state drive (SSD), a USB flash drive, a CD, etc. In these examples, the deployment unit 42C enables other devices to run the trained model 46B in the execution phase 16 to generate passive structure designs according to the reverse design-based techniques of the present disclosure.

[0068] Figure 4A and Figure 4BSeparate examples illustrate three-dimensional representation 54A and two-dimensional representation 54B of a dielectric lens generated by a GAN-based embodiment of the trained model 46B when run during execution phase 16. That is, in the use case scenarios of Figure 4A and Figure 4B , the trained model 46B represents an adversarially trained generator network that the training unit 42A has determined has achieved convergence. The trained model 46B implements the reverse design techniques of the present disclosure to generate the dielectric lens designs shown in Figure 4A and Figure 4B . Figure 4A The three-dimensional representation 54A of

[0069] illustrates two different protrusion regions, namely, the neck ring 53 and the island 55. The height (or amount of protrusion) of the neck ring 53 and the island 55 is shown by their relationship to the base of the dielectric lens in the two-dimensional representation 54B. Figure 4A and Figure 4B , and in the case of Figure 5 described below, a portion of the expected performance metric 44D is input to the execution phase 16, and the illustrated lens designs are generated to provide the input portion of the expected performance metric 44D when the corresponding lens is incorporated into a phased antenna array. In the specific cases of

[0070] Figure 5 , and in the case of Figure 4A and Figure 4B described below, the input portion of the expected performance metric 44D indicates a desired 1 dB increase in gain at the boresight (zero-degree) scan angle. Figure 5 The results illustrated are generated using modeling and simulation techniques relative to one of the lens designs generated by the trained model 46B. The gain values provided by the generated lens designs (per the modeling and simulation) are shown by the modeled gain 56. The expected gain 58 plots the gain value specified in the portion of the expected performance metric 44D provided as input to the trained model 46B during the execution phase 16.

[0071] The reference line 62 illustrates the gain provided by a phased antenna array not equipped with a dielectric lens designed according to any of the output passive structure designs 44E. As shown in the graph 60, when compared to the reference line 62 (which illustrates the gain provided by an antenna array not equipped with the above dielectric lens), the modeled gain 56 (provided by a phased antenna array equipped with a dielectric lens conforming to one of the output passive structures 44E) provides an average increase of 0.96 dB in the scan angle range of [0, 20] degrees and an average increase of 0.77 dB in the scan angle range of [0, 30] degrees.

[0072] Figure 6A and Figure 6B respectively illustrate three - dimensional representation 64A and two - dimensional representation 64B of a dielectric lens generated by a GAN - based embodiment of the trained model 46B when running in execution stage 16. Based on the input parameters of the expected performance metric 44B corresponding to an expected 2dB increase in gain at the boresight (zero - degree) scan angle, the training model 46B implements the inverse - design technique of the present disclosure to generate Figure 6A and Figure 6B the dielectric lens designs shown. Figure 6A The three - dimensional representation 64A of [[ ]] illustrates two different protrusion regions, namely, the neck ring 63 and the island 65. The height (or the amount of protrusion) of the neck ring 63 and the island 65 is shown by their relationship to the base of the dielectric lens in the two - dimensional representation 64B.

[0073] Figure 7 is a graph 70 illustrating the gain values provided by a phased - array antenna incorporating a dielectric lens that combines GAN - generated designs shown in [[ ]] and [[ ]]. Figure 6A and Figure 6B The results illustrated are generated using modeling and simulation techniques relative to one of the lens designs generated by the trained model 46B. The gain values provided by the generated lens designs (according to the modeling and simulation) are shown by the modeled gain 66. The expected gain 68 plots the gain values specified in a portion of the expected performance metric 44D provided as input to the trained model 46B during execution stage 16. Figure 7 The control line 72 illustrates the gain provided by a phased - array antenna not equipped with a dielectric lens designed according to any output passive - structure design 44E. As shown in graph 70, as the scan angle increases from zero at the boresight, the observed gain gradually decreases.

[0074]

[0075] Figure 8 Figure 8 is a graph 80 showing the increase in peak gain (or maximum gain) provided by a phased - array antenna equipped with a dielectric lens of the inverse - design of the present disclosure compared to the peak gain provided by other phased - array antennas. The inverse - design gain 74 plots the peak gain provided by a phased - array antenna incorporating a dielectric lens formed according to the inverse - design technique of the present disclosure (e.g., as generated by the trained model 46B during execution stage 16). As shown in [[ ]] and [[ ]], the inverse - design gain 74 is a relatively stable value that does not significantly gradually decrease until the scan angle moves significantly away from the boresight. In contrast, the control line 76 plots the peak gain provided by a phased - array antenna not equipped with any additional passive structure. shown, the inverse - design gain 74 is a relatively stable value that does not significantly gradually decrease until the scan angle moves significantly away from the boresight. In contrast, the control line 76 plots the peak gain provided by a phased - array antenna not equipped with any additional passive structure.

[0076] Compared to the inverse design gain 74, the control line 76 follows a similar trajectory but provides a significantly and consistently lower peak gain starting from the line of sight and continuing through the widest scan angle. The hemispherical lens gain 78 plots the peak gain provided by a phased antenna array equipped with a hemispherical lens. Compared to the inverse design gain 74 or the control line 76, the hemispherical lens gain 78 tapers off more steeply and starts at a significantly narrower scan angle. Thus, the graph 80 shows that a dielectric lens designed according to the inverse design technique of the present disclosure improves the performance (in terms of peak gain) of the phased antenna array into which it is incorporated and provides an excellent performance improvement (in terms of peak gain) when compared to phased antenna arrays equipped with other types of passive structures.

[0077] Figure 9 Graph 90 is a graph comparing the increase in peak gain (or maximum gain) provided by phased antenna arrays equipped with modified and unmodified versions of the dielectric lens of the inverse design of the present disclosure and by phased antenna arrays not enhanced by any dielectric lens infrastructure. The full lens gain line 82 plots the peak gain provided by a phased antenna array incorporating a dielectric lens formed according to the inverse design technique of the present disclosure as the scan angle moves away from the line of sight (i.e., starting from zero and increasing along the x - axis of graph 90), e.g., as generated by the trained model 46B during execution phase 16.

[0078] The control line 84 plots the peak gain provided by a phased antenna array not enhanced by any dielectric lens infrastructure. The island - only lens line 86 plots the peak gain provided by a phased antenna array enhanced by a modified version of the dielectric lens corresponding to the full lens gain line 82. In the case of the island - only lens line 86, the phased antenna array is equipped with a dielectric lens in which the neck ring (e.g., neck ring 53 or neck ring 63) is removed and the island (e.g., island 55 or island 65) is the only remaining protrusion.

[0079] The neck - ring - only lens line 88 plots the peak gain provided by a phased antenna array enhanced by another modified version of the dielectric lens corresponding to the full lens gain line 82. In the case of the island - only lens line 86, the phased antenna array is equipped with a dielectric lens in which the island (e.g., island 55 or island 65) is removed and the neck ring (e.g., neck ring 53 or neck ring 63) is the only remaining protrusion.

[0080] By comparison with the performance plotted by the reference line 84, only the island lens line 86 shows the peak gain at the improved visual axis of the islands 55 / 65, and only the collar lens line 88 shows the collar 53 / 63 improving the peak gain as the scan angle moves away from the zero-degree scan at the visual axis. Thus, the combination of the visual axis shown by the full lens gain line 82 and the improvement through widening the scan angle exemplifies that the inverse-designed dielectric lens of the present disclosure (including both the collar 53 / 63 and the islands 55 / 65) provides an overall peak gain improvement at all scan angles, as shown by comparison with the reference line 84.

[0081] Figure 10 is a block diagram illustrating aspects of the training phase 14 with respect to the training model 46A. As part of the training phase 14, one or more of the training data pairs 44C are provided to the training model 46A (illustrated as the "generator network 46A" in Figure 10 ). The discriminator network 92 is trained using one or more engineered passive structure designs 94. The engineered passive structure designs 94 represent dielectric lens designs that, when integrated into a phased antenna array, produce significant performance improvements in simulation or field testing. Thus, the engineered passive structure designs 94 can be regarded as the ground truth data for training the discriminator network of the GAN, in which case the discriminator network is represented by the discriminator network 92.

[0082] The discriminator network 92 performs adversarial training with respect to the generator network 46A by outputting predictions 96. Each of the generator network 46A and the discriminator network 92 can represent a deep neural network. The prediction 96 represents a probability value that indicates whether a particular one of the output passive structure designs 44E output by the generator network 46A during a training iteration will produce a portion of the expected performance metric 44D when integrated into a phased antenna array. The iterations of the training phase 14 iterate to improve the data accuracy delivered by the generator network 46A and, in some instances, the prediction accuracy delivered by the discriminator network 92.

[0083] After several iterative runs of the training phase 14 and based on the value of the prediction 96 output by the discriminator network 92, the training unit 42A can determine that the generator network 46A has achieved convergence. At this time, the training unit 42A can discard the discriminator network 92 and re-characterize the generator network 46A as the trained model 46B. The inverse design aspect of the present disclosure is represented by the functionality described with respect to the generator network 46A (whether during training or after convergence), because the generator network 46A outputs a dielectric lens design based on the expected performance information.

[0084] Figure 11is a block diagram example of an autoencoder neural network (ANN)-based architecture of an exemplary training model 46A and a trained model 46B. ANN 100 includes an encoder network 102 and a decoder network 104. According to some embodiments, each of the encoder network 102 and the decoder network 104 represents a separate neural network. Particularly when they are jointly trained, the encoder network 102 and the decoder network 104 can implement an efficient way to encode a large set of different images into a relatively small (n-dimensional) vector space, which is Figure 11 illustrated by the latent space representation 106 in

[0085] By exploring the representations in the latent space via one or more of perturbation, interpolation, or extrapolation methods, the ANN-driven inverse design techniques of the present disclosure can generate customized passive structure designs that improve the performance of the phased antenna arrays into which they are integrated. Some techniques of the present disclosure utilize these properties of the ANN to achieve inverse design generation and optimization of passive structure designs.

[0086] The encoder network 102 receives input data 108. The encoder network 102 performs dimensionality reduction and other processing operations on the input data 108 to form the latent space representation 106. The decoder network 104 takes the latent space representation 106 as input and reconstructs the latent space representation 106 as close as possible to the input data 108 to form output data 98. In the scenario of ideal or perfect performance of ANN 100, the output data 98 would exactly match the input data 108. In other words, in the best scenario, the decoder network 104 would reconstruct the latent space representation losslessly. In sub-optimal use case scenarios, the output data 98 differs from the input data 108 by an increment called "loss" or "reconstruction loss". The accuracy / precision of the reconstruction of the output data 98 compared to the input data 108 is represented by a "reconstruction quality function".

[0087] The encoder network 102 and / or the decoder network 104 may include multiple categories of predictive deep learning or machine learning models that work in concert, can be them, or can be part of them. As part of the training phase 14, the training unit 42A can generate a dataset including a number (e.g., approximately thousands) of candidate dielectric lens designs. One of the goals of the training phase 14 is to train ANN 100 to learn to generate the latent space representation 106 in a way that minimizes the loss in the full lens reconstruction represented in the output 98. The training unit 42A attempts to minimize the loss through the training of both the encoder network 102 and the decoder network 104.

[0088] With each iteration of training phase 14, ANN 100 improves its ability to reconstruct the output passive structure design 44E in the design space (which represents the output data 98) from the input data 108 after passing through the n-dimensional latent space representation 106, which acts as a bottleneck in the inverse design process. The training unit 42A can compute a loss function that compares the output data 98 (in the form of one of the output passive structure designs 44E) with the original passive structure design of the input data 108. The training unit 42A can express these losses as any type of relevant loss function, including but not limited to entropy, L1, L2, cosine similarity, or variants thereof.

[0089] The training unit 42A can use the derivative of the loss function with respect to the model parameters (e.g., the first derivative) in order to improve the model parameters using any one of the relevant gradient-based or gradient-free optimization techniques such as SGD, ADAM, ADAGRAD, RMSprop, BFGS, or others. The training unit 42A can iterate these steps as part of training phase 14 until the 42A resources are exhausted or the first occurrence of convergence of ANN 100 is achieved. The training unit 42A can utilize various types of computing resources to implement training phase 14, such as one or more of cloud-based computing resources, or various components of the processing circuitry 31 (e.g., CPU hardware, GPU hardware, etc.) in order to improve data accuracy with respect to execution phase 16.

[0090] After the training unit 42A concludes training phase 14, the design unit 42B can execute ANN 100 (which now represents the trained model 46B) or the deployment unit 42DC can deploy ANN 100 (which now represents the trained model 46B) to another device. In execution phase 16, ANN 100 can generate an instance of the latent space representation 106 for one or more passive structure designs 44E. The design unit 42B or a remote device executing execution phase 16 can perturb and decode a single instance of the latent space representation 106 to form one or more output passive structure designs 44E that are modified versions of the starting regular-shaped lens. The design unit 42B or a remote device executing execution phase 16 can interpolate / extrapolate and decode multiple instances of the latent space representation 106 to combine various features extracted from different dielectric lens designs in order to form a hybrid dielectric lens design as part of the output passive structure design 44E.

[0091] Figure 12Illustrates an example of a dielectric lens design generated by an ANN-based implementation of the trained model 46B. During execution phase 16, the ANN-based implementation of the trained model 46B can take the hemispherical lens design 110A as an execution phase input. The hemispherical lens design 110A can represent an engineered design. The ANN-based implementation of the trained model 46B can randomly perturb and decode a single instance of the latent space representation 106 to produce irregular lens shapes 110B, 110C, and 110D. Each of the irregular lens shapes 110B, 110C, and 110D represents a variation derived from the hemispherical lens shape 110A, which the ANN-based implementation of the trained model 46B can form via perturbation and decoding based on one or more of the performance metrics 44D.

[0092] Figure 13 Illustrates a top view of an example of a dielectric lens design generated by an ANN-based implementation of the trained model 46B. During execution phase 16, the ANN-based implementation of the trained model 46B can take the starting lens designs 112A and 112D as execution phase inputs. Using the starting lens designs 112A and 112D and one or more of the performance metrics 44D as execution phase inputs, the ANN-based implementation of the trained model 46B can interpolate across multiple instances of the latent space representation 106 (e.g., the corresponding n-dimensional representations of each of the starting lens designs 112A and 112D) to provide hybrid lens designs 112B and 112C. The ANN-based implementation of the trained model 46B can provide additional hybrid or intermediate designs as part of providing a smooth transition from the starting lens design 112A to the starting lens design 112D, and vice versa.

[0093] Figure 14A Is a scatter plot 114 that clusters fully encoded instances of the latent space representation 106 of various lens designs. The scatter plot 114 represents the clustering of the fully encoded latent space representations of 60,000 unique lens designs. The scatter plot 114 includes six different clusters, with each cluster representing a corresponding family of shape designs. Figure 14B Is an illustration Figure 14A Of a conceptual top view of the selection of the lens design represented in

[0094] Figure 14B Illustrates sampling of lens designs across the latent space representation 106, which shows the wide variety of newly discovered lens shape options produced by the ANN-based implementation of the trained model 46B.

[0095] The inverse design technique of dielectric lenses provides various technical improvements in the technical field of phased antenna array construction. As an example, the inverse design technique of the present disclosure reduces the time and resources that would otherwise be spent in implementing a dielectric lens design optimized for a specific use case. Otherwise, the time required to iterate through a traditional design optimization process is typically on the order of several years. In contrast, the inverse design technique of the present disclosure can be regarded as a series of three stages, namely data collection, training stage 14, and deployment.

[0096] The time taken for data collection depends on the application and can be as low as on the order of several minutes to on the order of several months. In many cases, the time for training is on the order of several hours and can vary based on the neural network (e.g., ANN) architecture. The time for deploying the generated design (e.g., output passive structure design 44E) is on the order of several seconds. Thus, overall, the inverse design technique of the present disclosure reduces the time taken for dielectric lens design generation and deployment from several years to at most several months.

[0097] As another example of the technical improvements provided by the inverse design technique of the present disclosure, the inverse design technique of the present disclosure provides a customized design that is suitable in terms of the scenario for the phased antenna array into which it is integrated. The various dielectric lens designs (e.g., Figure 4A , Figure 4B , Figure 6A , Figure 6B , Figure 12 and Figure 13 shown in ) do not conform to regular geometries. Instead, irregular surface curvatures and features are outside of conventional engineering design paradigms and represent lenses of irregular shapes generated using the inverse design technique of the present disclosure. As another example of the technical improvements provided by the system of the present disclosure, the neural network (e.g., GAN, ANN, etc.) architecture of the present disclosure tends to achieve convergence to a locally optimized design in endless cases, while traditional (e.g., Edisonian) design processes do not tend to this type of convergence in many scenarios.

[0098] Figure 15 illustrates the results of the GAN-based inverse design optimization technique of the present disclosure. The passive structure generation device 30 can utilize the GAN-based functionality to perform design optimization to generate Figure 15One or more of the results shown. Although the GAN-based inverse design optimization techniques of the present disclosure are mainly described in the context of the inverse design optimization of dielectric lenses as an example, it should be understood that the inverse design optimization techniques of the present disclosure can also be used in other application fields. The passive structure generation device 30 can implement the GAN-based inverse design optimization techniques of the present disclosure to form an optimal lens design (as determined using the expected performance metric 44D) starting from an inventory lens design, or an available lens design that is closest to the optimal lens design.

[0099] In various inverse design optimization specific implementations of the AI / ML model 46, the generator network and the discriminator network of the training model 46A can be any one of multiple categories of predictive deep learning or machine learning models. For example, the generator network can include the encoder and decoder networks of an autoencoder or a VAE. In other examples, the generator network can be a U-NET or can be composed of multiple U-NETs connected in series or in parallel. In some examples, the decoder of such a generator network can be a locked decoder from an autoencoder pre-trained in an unsupervised manner, or a β-VAE.

[0100] The discriminator network of the training model 46A can also be any member of a collection of deep learning or machine learning models (such as a deep convolutional neural network). Various examples of the GAN architecture according to the inverse design optimization techniques of the present disclosure include Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN), PixelGAN, CycleGAN, or any variants thereof. In some examples of the techniques of the present disclosure, the passive structure generation device 30 can use a variational autoencoder (VAE) in an independent manner or in combination with a GAN to perform the inverse design optimization techniques of the present disclosure. In some examples, residual connections and / or attention mechanisms can be incorporated into the generator and / or discriminator networks of the GAN represented by the training model 46A.

[0101] The training unit 42A can use the training data pair 44C to train the GAN-based inverse design optimizer specific implementation of the training model 46A. For example, the training unit 42A can train the generator to learn the inverse mapping from the expected performance part of the training pair 44C to the corresponding optimized lens design. In each iteration of the training phase 14, the generator network generates an optimized lens design given a specific part of the expected performance metric 44D as input (as part of one or more of the training pairs 44C). The training unit 42A can calculate the loss function by comparing the optimized design and the engineered design and by comparing the predictions of the discriminator for the given optimized design.

[0102] These operations are part of a conditional embodiment of the GAN-based optimizer of the present disclosure, where the optimization design is conditioned on having similarity or shared characteristics (or characteristics derived from those of the engineered design) with the engineered design. In each iteration of the discriminator training phase 14, the training unit 42A may calculate the loss function by comparing the prediction given to the optimization design with all-zero values and the prediction given to the engineered design with all-one values. To calculate the loss (as a result of the above comparison), the training unit 42A may use any relevant loss function (e.g., cross-entropy, L1, L2, cosine similarity, etc.) or a variant thereof.

[0103] The training unit 42A may use the derivative of the loss function with respect to the model parameters in order to improve or further refine the model parameters using any one of the relevant gradient-based or gradient-free optimization techniques (e.g., SGD, ADAM, ADAGRAD, RMSprop, BFGS, etc.). The training unit 42A may iterate these steps until convergence of the generator network is detected or until a state of resource exhaustion is determined, whichever occurs first. After the end of the training phase 14, the design unit 42B or the remote device may place the trained generator network (represented by the trained model 42B) in the execution phase 16 to output one or more optimized lens designs corresponding to those of the expected performance metric 44D provided as an input to the execution phase. In various use case scenarios, the design unit 42B or the remote device may output an optimized design in a single prediction iteration (e.g., by only outputting a desired design that meets or exceeds the expected performance metric 44D provided as an input to the execution phase), or may output multiple optimized designs in different iterations of the execution phase 16. In the latter scenario, the design unit 42B or the remote device may perform post-processing (e.g., by filtering multiple designs) to identify the best-performing design.

[0104] The hemispherical lens design 116 is an engineered design, such as a lens design manually generated by a research engineer. The irregular-shaped lens designs 118, 120, and 122 represent optimized lens designs that the design unit 42B or the remote device executing the trained model 46B may output using the hemispherical lens design 116 as a starting point. Each of the irregular-shaped lens designs 118, 120, and 122 represents an optimized design formed by the trained model 46B from the hemispherical lens design 116 (either in a single step or through a series of optimization steps) based on a part of the expected performance metric 44D provided as an input to the execution phase. By adopting the expected performance metric 44D and generating one or more of the irregular-shaped lens designs 118, 120, or 122 as output, the trained model 46B performs the reverse design optimization technique of the present disclosure.

[0105] According to some aspects of the present disclosure, the training unit 42A can train the training model 46A to improve the fidelity of the lens design generated and / or optimized using the various deep learning methods described above. As used herein, the term "fidelity" refers to a measure of the similarity between the generated / optimized lens design output by the training model 46A or the trained model 46B and a conventional engineered lens design regarded as the ground truth lens design.

[0106] In one example, the training unit 42A can first train an autoencoder network in an unsupervised manner. Subsequently, the training unit 42A can use the pre-trained decoder of the autoencoder as the locked decoder in the generator of the GAN represented by the training model 46A. In this way, the training unit 42A can substantially use the pre-trained decoder with locked weights as part of the generator network of the GAN represented by the training model 46A. In this way, the training unit 42A can reduce the complexity of the learning task of the generator network. The training unit 42A trains the generator network to learn the mapping from the expected performance metric 44D to the characterized representation of the lens design, rather than requiring the generator network to learn the mapping from the expected performance metric 44D to the full lens design. Generally, the dimension of the characterized representation is one or more orders of magnitude lower than the dimension of the full representation of the same lens design.

[0107] In experiments, the specific implementation of the GAN-based generator of the trained model 46B improves the fidelity of the output passive structure design 44E. Similarly, the improved fidelity of the output passive structure design 44E indicates that the output passive structure design 44E is more closely similar to the engineered lens design used as the ground truth design. These specific implementations of the trained model 46B provide a technical improvement in imposing constraints (such as manufacturability constraints) on the output passive structure design 44E.

[0108] In some examples, the generator network of the GAN-based implementations of the training model 46A and the trained model 46B is a U-NET. The U-NET can be composed of an encoder and a decoder (where each of the encoder and the decoder is composed of multiple convolutional blocks, non-linearities, downsampling or upsampling operations, and transposed convolutions). In some implementations consistent with aspects of the present disclosure, the U-NET representing the generator network may not include residual or skip connections (e.g., res-net type connections), while in other implementations consistent with aspects of the present disclosure, the U-NET representing the generator network may not include residual or skip connections (e.g., res-net type connections).

[0109] At the first stage, the training unit 42A may use a dataset consisting of designs to train the first U-NET. The training unit 42A may pass through the design such that the U-NET receives an encoded representation and decodes the encoded representation to form a resulting output. The training unit 42A may compare the resulting output of the U-NET with the original design. At this stage, the first U-NET learns an identity mapping and encodes the design into a latent space representation. Subsequently, the training unit 42A trains the decoder portion of the first U-NET to learn the mapping from the latent representation to the design.

[0110] At the second stage, as a benchmark for comparison, the training unit 42A may train a second U-NET to learn the mapping from performance to design. Thus, the training unit 42A may use pairs or tuples having a (performance, design) structure to train the second U-NET. At the third stage, the passive structure generation device 30 may construct a third U-NET using an architecture similar to the first two U-NETs, where the decoder of the first UNET in an immutable (or non-modifiable) form is used as the decoder in the third U-NET. The training unit 42A may use (performance, design) tuples to train the third U-NET.

[0111] In some examples, the passive structure generation device 30 may maintain corresponding or sometimes even identical architectures and training hyperparameters (e.g., learning rate, loss function, number of epochs, training set and validation set sizes, etc.) for all three of the above U-NETs so as to keep the experiments as comparable as possible. A non-limiting example of a loss function that the training unit 42A may use is binary cross-entropy, such as in the case where the synthetic dataset consists of images with binary pixel values for lens design. In some examples consistent with aspects of the present disclosure, the training unit 42A may implement the pre-trained decoder in a variable (or modifiable) form. In these examples, the training unit 42A provides a “warm start” to the pre-trained decoder with respect to the third U-NET because the weights of the decoder of the third U-NET are not randomly initiated but are set to the weights of the decoder of the first U-NET.

[0112] Experimental results show that the pre-trained decoder reduces the training and validation loss values and produces a higher fidelity with respect to the lens designs generated during the execution phase 16. These experimental results reflect the fidelity of the original output from the S-shaped layer at the end of the U-NET chain, even without the benefit of thresholding or other post-processing. While the experimental results of the specific implementation of the U-NET chain of the present disclosure are for higher fidelity lens designs for a particular application (in this case, for integration into a 5G phased antenna array), it should be understood that the U-NET chain-based techniques of the present disclosure may also be applied to improve the fidelity of inverse automated designs generated for other end uses.

[0113] Some aspects of the present disclosure relate to systems and techniques for generating a data set represented by a training data pair 44C. In these examples, a passive structure generation device 30 or another device configured according to these aspects of the present disclosure can generate a labeled data set represented by the training data pair 44C, which enables supervised learning-based training of a deep learning model (e.g., GAN) represented by a training model 46A. Similarly, the labeled data set represented by the training data pair 44C includes 2-tuples having a (performance, design) structure.

[0114] A device configured to implement the labeled data set generation techniques of the present disclosure can use EM simulation or measurement to characterize the performance of a finite set of lens designs. In some examples, the device can select a pair of these lens designs (hereinafter, "parent designs") from the finite set and interpolate the two parent designs to generate one or more "offspring designs". For example, the device can use an autoencoder pre-trained in an unsupervised manner to interpolate between two parent designs to generate offspring designs.

[0115] The device can use the interpolation of the performance attributed to the parent designs to simulate / estimate the performance of the offspring designs. In this way, the device can generate new 2-tuples having an (offspring design, estimated performance) structure, which can be used to populate the training pair 44C. The device can select two parent designs from a set of N parent designs in a combinatorial number of ways and generate dozens of interpolated offspring designs from each corresponding parent design pair. In this way, the devices of the present disclosure can provide a technical improvement in increasing the data set size (also referred to as "data augmentation") with respect to the training data pair 44C.

[0116] By providing data augmentation with respect to the training data pair 44C, the devices of the present disclosure provide a technical improvement in the improved data accuracy of the prediction functionality of the trained model 46B. Among the technical improvements provided by these techniques of the present disclosure, the data accuracy is improved through performance prediction / estimation. By predicting the estimated performance of the offspring lens designs, these techniques of the present disclosure provide data augmentation with respect to the synthetic data set used to generate the training data pair 44C, thereby improving the accuracy of the training phase 14 with respect to the training module 46A.

[0117] Figure 16 Examples of interpolated lens designs generated as part of the labeled data set generation techniques of the present disclosure are illustrated. In Figure 16 the example, parent designs 124A and 124B represent two different engineered designs that can be used as starting points at both ends by the neural network (e.g., ANN) of the present disclosure during the interpolation process. The systems of the present disclosure can perform simulation techniques (e.g., EM solver) to simulate the performance. The labeled data set generation neural network of the present disclosure can interpolate between the parent designs 124 to generate several hybrid lens designs, three of which are shown inFigure 16 are illustrated as descendant designs 126A-C in

[0118] In some examples, the interpolation is based on the assumption that the parent designs 124 represent upper and lower limits in terms of lens performance, and that the performance of each of the descendant designs 126 (as an interpolation of the pair of parent designs 124) can be estimated by interpolating the performance of the parent designs 126 via use of the same weighted average formula used to generate each respective descendant design 124. This assumption can be expressed in the form of the following equations (1) and (2):

[0119] If

[0120] lens_offspring = alpha * lens_parent_first+(1 - alpha)*lens_parent_second

[0121] …(1)

[0122] Then

[0123] perf_offspring = alpha * perf_parent_first+(1 - alpha)*perf_parent_second

[0124] …(2)

[0125] where "perf" represents performance. Different values of the constant represented by "alpha" in equation (2) above are used to generate the corresponding performance estimates for each of the descendant designs 126.

[0126] The process of multiple experiments confirmed the assumptions described with respect to equations (1) and (2), as the simulated performance of the interpolated lenses described by the descendant designs 126 matched or was relatively close to the interpolated performance metric obtained using equation (2). In one example, the performance of parent design 124A was quantified as a figure of merit (FoM) of -0.5, while the FoM of parent design 124B was 1.86. A labeled dataset generation neural network (ANN in this experiment) interpolated the parent designs 124 to generate the descendant designs 126, such that different values of "alpha" in equation (1) were used to generate each of the descendant designs 126A, 126B, and 126C.

[0127] In one experimental example, the master design 124 represents one of a pair of master designs out of 3,240 pairs formed from 81 candidate master lens designs. As part of this particular experiment, the labeled dataset of the present disclosure generates ANNs to interpolate each corresponding pair of master designs in twenty ways, resulting in a total of 64,800 data points. Compared with existing EM modeling-based techniques that would take over seven years to generate 64,800 data points, the ANNs of the present disclosure generate 64,800 data points in less than a week.

[0128] Accordingly, when compared with the prior art, the system of the present disclosure provides a technology enhancement with data augmentation at a significantly reduced time and a significantly reduced resource expenditure. While the experiments described herein demonstrate technology improvements (e.g., reduced data precision and resource occupancy) for a specific application of synthetic lens design relative to 5G phased antenna arrays as a non-limiting example, it should be understood that the dataset augmentation technology of the present disclosure is also applicable to various other uses of inverse automated design solutions.

[0129] In the detailed description of exemplary embodiments of the present invention, reference is made to the accompanying drawings, which illustrate specific embodiments in which the present invention may be practiced. The exemplary embodiments are not intended to enumerate all embodiments in accordance with the present 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 present invention. Accordingly, the following detailed description is not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims.

[0130] Unless otherwise specified, all numbers expressing feature sizes, amounts, and physical properties used in this specification and the claims are to be understood to be modified in all instances by the term "about," "approximately," or "substantially." Accordingly, unless indicated to the contrary, the numerical parameters set forth in the foregoing specification and attached claims are approximations that may vary depending upon the desired properties sought to be obtained by those skilled in the art utilizing the teachings disclosed herein.

[0131] Unless the context clearly dictates otherwise, as used in the specification and the appended claims, the singular forms "a," "an," and "the" encompass embodiments having plural referents. Unless the context clearly dictates otherwise, as used in the specification and the appended claims, the term "or" is generally employed in its sense including "and / or."

[0132] It should be recognized that, according to this example, certain actions or events of any of the methods described herein may be implemented in a different order, may be added together, combined, or omitted (e.g., not all of the described actions or events are necessary for the practice of the method). Additionally, in some examples, the actions or events may be performed, for example, through multithreading, interrupt processing, or by multiple processors simultaneously rather than sequentially.

[0133] The techniques described in this disclosure may be implemented, at least in part, in hardware, software, firmware, or any combination thereof. For example, aspects of the described techniques may 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 circuitry, as well as any combination of such components. The term "processor" or "processing circuitry" generally may refer to any of the foregoing logic circuitry alone or in combination with other logic circuitry, or any other equivalent circuitry. A control unit that includes hardware may also perform one or more of the techniques of this disclosure.

[0134] 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. Additionally, any of the described units, modules, or components may be implemented together or separately as discrete but cooperating logic devices. Depicting 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 functions associated with one or more modules or units may be performed by separate hardware or software components or integrated within common or separate hardware or software components.

[0135] The techniques described in this disclosure may also be embodied or encoded in a computer-readable medium, such as a computer-readable storage medium, that includes instructions. For example, when executed, the instructions embedded or encoded in the computer-readable storage medium may cause a programmable processor or other processor to perform the method. The computer-readable storage medium 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, a hard disk, a CD-ROM, a floppy disk, a cassette tape, magnetic media, optical media, or other computer system-readable media.

[0136] Various examples have been described. These examples, as well as other examples, are within the scope of the following claims.

Claims

1. A method for generating a passive structure design, the method comprising: Receiving, by a processing circuit of a computing device, a set of performance metrics; Providing, by the processing circuit, the set of performance metrics to a trained neural network; Receiving, by the processing circuit, from the trained neural network a passive structure design associated with a passive structure; and Outputting, by the processing circuit via a communication interface, the passive structure design.

2. The method according to claim 1, the method further comprising: Generating, by the processing circuit, a plurality of candidate passive structure designs based on the passive structure design; And Selecting, by the processing circuit, a final passive structure design from the plurality of candidate passive structure designs as the output passive structure design.

3. The method according to claim 2, wherein generating the plurality of candidate passive structures comprises sequentially modifying the passive structure design by the processing circuit in multiple iterations.

4. The method according to claim 3, wherein sequentially modifying the passive structure design in the multiple iterations comprises randomly modifying a latent space representation of the passive structure, each latent space representation being associated with a corresponding candidate passive structure design among the plurality of candidate structure designs.

5. The method according to claim 2, wherein selecting the final passive structure design comprises: Providing, by the processing circuit, each candidate passive structure design included in the candidate passive structure designs to a simulator; And Receiving, by the processing circuit, from the simulator at least one performance metric associated with each candidate passive structure design.

6. A method for training a neural network, the method comprising: Training a first neural network to generate a latent space representation based on a passive structure design, the first neural network comprising a first encoder and a first decoder; And Training a second neural network to generate an output passive structure design based on performance metrics, the second neural network comprising a second encoder and a second decoder, the second neural network comprising the first decoder.

7. The method according to claim 6, wherein during the training of the second neural network, the weights included in the second decoder are in an immutable form.

8. A passive structure design apparatus, comprising: Components for receiving a set of performance metrics; Components for providing the set of performance metrics to a trained neural network; Components for receiving from the trained neural network a passive structure design associated with a passive structure; And Components for outputting the passive structure design.

9. A neural network training apparatus, comprising: Components for training a first neural network to generate a latent space representation based on a passive structure design, the first neural network comprising a first encoder and a first decoder; And Components for training a second neural network to generate an output passive structure design based on performance metrics, the second neural network comprising a second encoder and a second decoder, the second neural network comprising the first decoder.

10. A passive structure generation device, comprising: At least one non-transitory computer-readable storage medium having instructions stored thereon; At least one processor, the at least one processor being coupled to the at least one non-transitory computer-readable storage medium and configured to execute the instructions to: Receive a set of performance metrics; Provide the set of performance metrics to a trained neural network; Receive a passive structure design from the trained neural network, the passive structure design being associated with a passive structure; and Output the 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.

11. The passive structure generation device according to claim 10, wherein the processor is further configured to: Generate a candidate passive structure design based on the passive structure design; Select a final passive structure design from the candidate passive structure designs as the output passive structure.

12. The passive structure generation device according to claim 11, wherein generating the candidate passive structure design includes sequentially modifying the passive structure design in multiple iterations.

13. The passive structure generation device according to claim 12, wherein sequentially modifying the candidate passive structure design includes randomly modifying the latent space representation associated with each candidate passive structure design.

14. The passive structure generation device according to claim 11, wherein selecting the final passive structure design includes: Providing each candidate passive structure design included in the candidate passive structure designs to a simulation technique; And Receiving at least one performance metric associated with each candidate passive structure design from the simulation technique.

15. The passive structure generation device according to claim 10, wherein the trained neural network was previously trained by: Training a first neural network to generate a latent space representation based on a passive structure design, the first neural network including a first encoder and a first decoder; Training a second neural network to generate a passive structure design based on performance metrics, the second neural network including a second encoder and a second decoder, the second neural network including the first decoder.

16. The passive structure generation device according to claim 15, wherein during the training of the second neural network, the weights included in the second decoder are locked.

Images