WO2020185207A1

WO2020185207A1 - Computerized system and method for generative circuit design with machine-learned networks

Info

Publication number: WO2020185207A1
Application number: PCT/US2019/021628
Authority: WO
Inventors: Krzysztof CHALUPKA; Craig BEEBE; James Donnelly; Darrell A. Teegarden; Janani VENUGOPALAN; Sanjeev SRIVASTAVA; Lucia MIRABELLA; Suraj Ravi MUSUVATHY
Original assignee: Siemens Aktiengesellschaft; Siemens Energy, Inc.
Priority date: 2019-03-11
Filing date: 2019-03-11
Publication date: 2020-09-17

Abstract

Computerized system and method for generative circuit design are provided. The method includes estimating, by way of an artificial intelligence processor (12), sets of design parameters likely to meet desired circuit requirements of a circuit to be designed. The estimating by a generative machine-learned network (14) in response to input of the desired circuit requirements. The method further includes predicting values of key performance indicators (KPIs). The predicting by a forward predictive machine-learned network (16) in response to input of the sets of design parameters. The KPIs indicative of respective performance levels of the sets of estimated design parameters relative to the desired circuit requirements. Based on the predicted values of the KPIs, evaluating whether at least one of the sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements, and selecting a set of design parameters from the sets of estimated design parameters that meet the desired circuit requirements of the circuit to be designed.

Description

COMPUTERIZED SYSTEM AND METHOD FOR GENERATIVE

CIRCUIT DESIGN WITH MACHINE-LEARNED NETWORKS

[0001] BACKGROUND

[0002] 1. FIELD

[0003] Disclosed embodiments relate generally to the field of electrical circuit design, and, more particularly, to circuit design autonomously configured by way of machine-learned networks, and, even more particularly, to computerized system and method for generative design of circuit parameters chosen by way of the machine-learned networks to meet desired circuit requirements.

[0004] 2. Description of the Related Art

[0005] In the design of electrical circuits, a design engineer typically may face a

myriad of decisions based on circuit requirements and criteria,

which, for example, may dictate the specifications of any given circuit in terms of design constraints, etc. The design constraints may in turn limit the choice of design techniques, component selection, design parameters, etc. The design engineer has to satisfy the design constraints and, at the same time, try to optimize the design, for example, in terms of suitable design parameter combinations. This task becomes progressively difficult and burdensome as the complexity of the circuit increases. For example, the design engineer would have to perform time-consuming and burdensome trade-off analyses to attempt finding appropriate choices for a plurality of design parameters.

[0006] To try to alleviate some of the difficulties involved, simulation software

—such as SystemVision® simulation software, produced by Mentor, a

Siemens business— may be used to, for example, verify whether the circuit parameters hand-picked by the design engineer meet the desired requirements and criteria. [0007] At least in view of the foregoing consideration, for relatively large circuits or in case of subsystems involving a plurality of circuits, the design phase of a circuit or subsystem can be substantially complex and time-consuming. In addition, validating a large circuit design using simulation software can consume a substantial amount of time, and this can considerably slow down the typical design-evaluate-redesign iterative process that may be involved to eventually reach a final set of design parameters for a given circuit.

[0008] For examples of certain known approaches involving electronic design

automation (EDA) that purportedly leverage computation to try to optimize the design of circuitry, see paper by Thompson A., Layzell P., and Zebulum R. S., titled“Explorations In Design Space: Unconventional Electronics Design Through Artificial Evolution”, published in IEEE Transactions on

Evolutionary Computation, Vol. 3, Issue 3, Sept., 1999, pp.167-196; and also see paper by Vassilev V. K., Job D., and Miller J. F., titled“Towards the Automatic Design of More Efficient Digital Circuits”, published in

Proceedings of the 2nd NASA/DoD Workshop on Evolvable Hardware, IEEE, pp. 151-160. It will be appreciated that the foregoing approaches may be limited in broad and cost-effective industrial applicability since their efficacy may be contingent on the substantial involvement of highly technical personnel with great and specialized expertise in creating complex circuitry.

[0009] BRIEF DESCRIPTION

[0010] A disclosed embodiment is directed to a computerized method. Without

limitation, the method may include estimating, by way of an artificial intelligence processor (12), one or more sets of design parameters likely to meet desired circuit requirements of a circuit to be designed. The estimating may be by a generative machine-learned network (14) in response to input of the desired circuit requirements. The method may further include predicting, by way of artificial intelligence processor (12), values of one or more key performance indicators (KPIs). The predicting may be by a forward predictive machine-learned network (16) in response to input of the one or more sets of design parameters estimated by the generative machine-learned network. The predicted values of the one or more KPIs being indicative of respective performance levels of the one or more estimated sets of design parameters relative to the desired circuit requirements. Based on the predicted values of the one or more KPIs of the one or more sets of design parameters estimated by the generative machine-learned network, evaluating whether at least one of the one or more sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements; and selecting a set of design parameters from the at least one of the one or more sets of estimated design parameters that meets the desired circuit requirements of the circuit to be designed.

[0011] A further disclosed embodiment is directed to a computerized system.

Without limitation, an artificial intelligence processor (12) may be configured to estimate one or more sets of design parameters likely to meet desired circuit requirements of a circuit to be designed, where a generative machine-learned network (14) of the artificial intelligence processor may be configured to generate the estimates of the one or more sets of design parameters in response to input of the desired circuit requirements. The artificial intelligence processor may be further configured to predict values of one or more key performance indicators (KPIs), where a forward predictive machine-learned network (16) of the artificial intelligence processor may be configured to generate the predicted values of the one or more key KPIs in response to input of the one or more sets of design parameters estimated by the generative machine-learned network. The predicted values of the one or more KPIs being indicative of respective performance levels of the one or more estimates of sets of design parameters relative to the desired circuit requirements. An evaluator (18) may be configured to evaluate, based on the predicted values of the one or more KPIs, whether at least one of the one or more sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements. A set of design parameters is selected from the at least one of the one or more sets of design parameters estimated by the generative machine-learned network that meets the desired circuit

requirements of the circuit to be designed. [0012] BRIEF DESCRIPTION OF THE DRAWINGS

[0013] FIG. 1 illustrates a block diagram of one non-limiting embodiment of a

disclosed computerized system, as may be used for carrying out generative circuit design.

[0014] FIG. 2 is a flow chart of one non-limiting embodiment of a computerized

method for machine training disclosed networks.

[0015] FIG. 3 is a flow chart of one non-limiting embodiment of a disclosed

computerized method, as may be used for carrying out generative circuit design.

[0016] FIG. 4 illustrates one non-limiting embodiment of a circuit formed by a given arrangement (e.g., topology) of circuit components whose parameter values are to be generatively designed by way of disclosed embodiments.

[0017] FIG. 5 illustrates a non-limiting example of machine-learned estimation of sets of design parameters from circuit requirements.

[0018] FIG. 6 illustrates a non-limiting example of machine-learned prediction of respective values of KPIs from the sets of estimated design parameters.

[0019] FIG. 7 illustrate respective plots illustrating four non-limiting examples of possible sets of design parameters estimated by disclosed embodiments for the circuit illustrated in FIG. 4.

[0020] FIG. 8 are respective plots illustrating respective performances of the sets of estimated design parameters illustrated in FIG. 7 in term of deviations of predicted KPIs calculated by disclosed embodiments relative to simulated true values of circuit requirements. [0021] DETAILED DESCRIPTION

[0022] The present inventors have recognized that when a circuit involves a relatively large number of circuit components, or in situations involving a subsystem comprising a plurality of circuits, it may be difficult to efficiently and consistently optimize design of circuit parameters while still appropriately meeting any desired circuit or subsystem requirements. For example, as noted above, under current approaches, it is substantially challenging for a non expert to identify suitable parameter values or interrelationships between such parameters to achieve the desired circuit requirements.

[0023] At least in view of such recognition, disclosed embodiments formulate an innovative approach for cost-effectively and reliably generatively designing circuit parameters autonomously selected to appropriately meet the desired circuit requirements. Without limitation, disclosed embodiment may use the versatility of machine-learning methodologies to efficiently process vast quantities of data to generate and evaluate a multiplicity of possible circuit designs to achieve the desired circuit or subsystem requirements.

[0024] In the following detailed description, various specific details are set forth in order to provide a thorough understanding of such embodiments. However, those skilled in the art will understand that disclosed embodiments may be practiced without these specific details that the aspects of the present invention are not limited to the disclosed embodiments, and that aspects of the present invention may be practiced in a variety of alternative embodiments. In other instances, methods, procedures, and components, which would be well- understood by one skilled in the art have not been described in detail to avoid unnecessary and burdensome explanation. [0025] Furthermore, various operations may be described as multiple discrete steps performed in a manner that is helpful for understanding embodiments of the present invention. However, the order of description should not be construed as to imply that these operations need be performed in the order they are presented, nor that they are even order dependent, unless otherwise indicated. Moreover, repeated usage of the phrase“in one embodiment” does not necessarily refer to the same embodiment, although it may. It is noted that disclosed embodiments need not be construed as mutually exclusive embodiments, since aspects of such disclosed embodiments may be appropriately combined by one skilled in the art depending on the needs of a given application.

[0026] FIG. 1 illustrates a block diagram of one non-limiting embodiment of a

disclosed computerized system 10, as may be used for generative design of circuit parameters selected to satisfy desired circuit requirements of a circuit to be designed. In one non-limiting embodiment, a user-interface 20 may be used to enter a user input that, without limitation, may define circuit requirements 22 and evaluation criteria 24 of the circuit to be designed. Without limitation, circuit requirements may be defined in the form of at least one key

performance indicator (KPI) for the circuit to be designed. In one non limiting example, in a circuit 100 (FIG. 4) to be designed, such KPIs may be an amplification gain and a cutoff frequency. Without limitation, examples of evaluation criteria may include error tolerances defined by the user (or a specification) in connection with the circuit requirements, sensitivity of the desired circuit requirements to variation of one or more of the parameters in a given set of design parameters, cost or availability of components subject to a given set of design parameters, etc. [0027] It will be appreciated that disclosed embodiments may be applicable not just in connection with analog and/or digital circuits but may also be applied in the context of higher-level subsystems, such as may additionally involve sensors and actuators interacting with such circuits. Non-limiting example applications that may benefit from disclosed embodiments may include power electronics; power systems; motors & drives; robotics, sensors, and motion control signals and communications; electronics cooling and thermal control; lighting systems; electro-fluidic systems and control; electro-magnetic components; equipment involved in the generation of renewable energy, etc.

[0028] Desired circuit requirements 22 may be fed to a generative machine-learned network 14 that may be part of an artificial intelligence processor 12 configured to estimate one or more sets of design parameters (e.g., the parameter values of circuit components of circuit 100) likely to meet the desired circuit requirements (e.g., the desired amplification gain and cutoff frequency of circuit 100 (FIG. 4)). Generative machine-learned network 14 may be configured to generate the estimates of the one or more sets of design parameters (e.g., labelled in FIG. 1, as Set 1, Set 2 through Set n) in response to input of the desired circuit requirements.

[0029] In this context, we define generative network as any machine learning model which can take a set of requirements (e.g., circuit requirements) as input and give at least one set of design parameters as output. Non-limiting examples of generative methodologies that may be used in generative machine-learned network 14 may be Generative Adversarial Network (GAN) and Mixture Density Network (MDN). Based on the needs of a given application, generative machine-learned network 14 may be configured with the GAN methodology, the MDN methodology or both methodologies. [0030] For readers desirous of underlying theoretical principles in connection with said GAN methodology, reference is made to paper titled“Generative

Adversarial Nets” by Goodfellow I. J., Pouget-Abadie J., Mirza M., Xu B., Warde-Farley D., Ozair S., Courville A., and Bengio Y.; this paper is part of NIPS'14 Proceedings of the 27th International Conference on Neural

Information Processing Systems - Vol. 2, pp. 2672-2680, Dec. 2014, published by MIT Press Cambridge, MA, USA ©2014. Similarly, in connection with MDN methodology reference is made to paper titled“Mixture Density Network” by Bishop C. M., technical Report NCRG/94/004, Neural Computing Research Group, Aston University, Birmingham 1994.

[0031] In one non-limiting embodiment, artificial intelligence processor 12 is further configured to predict values of one or more key performance indicators (KPIs). Without limitation, a forward predictive machine-learned network 16 of artificial intelligence processor 12 may be configured to generate the predicted values of the one or more key KPIs in response to input of the one or more sets of estimated design parameters by generative machine-learned network 14. The predicted values of the one or more KPIs are indicative of respective performance levels of the one or more sets of estimated design parameters relative to the desired circuit requirements. In the foregoing example, the predicted values of the KPIs would indicate how the estimated design parameters of a given circuit design quantitatively track the desired gain and cutoff frequency.

[0032] Non-limiting examples of predictive methodologies that may be used in forward predictive machine-learned network 16 may involve deep residual learning, such as may be implemented in a fully-connected residual network. Without limitation, in lieu of performing a costly and time-consuming simulation, forward predictive machine-learned network 16 may be appropriately trained to predict an accurate approximation of simulation results practically in a fraction of a second for a given circuit, regardless of circuit size— as long as forward predictive network 16 is trained with an appropriate dataset. For readers desirous of underlying theoretical principles in connection with said predictive methodology, reference is made to paper titled “Deep Residual Learning for Image Recognition” by He K., Zhang X., Ren S., and Sun L, IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016.

[0033] In one non-limiting embodiment, an evaluator 18 may be configured to

evaluate, based on the predicted values of the one or more KPIs, whether at least one of the one or more sets of estimated design parameters by generative machine-learned network 14 meets the desired circuit requirements consistent with the evaluation criteria defined by the user. For example, whether error tolerances in connection with the circuit requirements and other criteria defined by the user have been fulfilled.

[0034] Once the desired circuit requirements have been appropriately fulfilled, a set of design parameters 26 that appropriately meets the desired circuit requirements may be selected. In the event, the evaluation by evaluator 18 determines that none of the one or more sets of design parameters estimated by generative machine-learned network 14 meets the desired circuit requirements, then generative machine-learned network 14 may be triggered to generate further estimates of sets of design parameters likely to meet the desired circuit requirements of the circuit to be designed. [0035] FIG. 2 is a flow chart of one non-limiting embodiment of a computerized method for machine training disclosed networks for carrying out generative circuit design. Without limitation, the training may use deep learning to respectively train generative machine-learned network 14 and forward predictive machine-learned network 16 (FIG. 1).

[0036] In block 30, training data, such as in connection with one or more topologies of a given circuit may be obtained. For machine training, hundreds, thousands, tens of thousands, or other number of samples with ground truth may be used. Without limitation, the training data may be obtained from simulation software, such as SystemVision® simulation software, produced by Mentor, a Siemens business.

[0037] In block 32, prior to the estimating by generative machine-learned network 14, a first dataset—such as may be obtained from respective simulations of respective topologies of the given circuit— may be used to train generative machine-learned network 14 as a generative model to learn estimating the one or more sets of design parameters corresponding to a given topology (of the simulated respective topologies) of the circuit to be designed.

[0038] Similarly, prior to the predicting by forward predictive machine-learned

network 14, a second dataset—obtained from respective simulations of respective topologies of the given circuit— may be configured to train forward predictive machine-learned network 14 as a forward predictive model to learn predicting the values of the one or more KPIs corresponding to the given topology of the circuit to be designed. Lastly, in block 34, the trained neural networks may be respectively stored for subsequent use in a given generative design. [0039] FIG. 3 is a flow chart of one non-limiting embodiment of a computerized method as may be used for carrying out generative circuit design in a disclosed artificial intelligence processor 12 (FIG. 1). Artificial intelligence processor 12 may be at least one processor configured for applying the foregoing machine-learned networks 14 and 16. Without limitation, artificial intelligence processor 12 may include massive parallel processing or may be a general processor, image processor, graphics processing unit, application specific integrated circuit, field programmable gate array, or other processor capable of using machine-learned networks.

[0040] In one non-limiting example, artificial intelligence processor 12 may be

configured to perform actions, such as depicted in block 44-48. A user or, alternatively, artificial intelligence processor 12 may be configured to performs actions, such as depicted in blocks 40 and 42. Once a design is selected, as depicted in block 50, a manufacturing system or assembly may perform actions depicted in block 52. Alternatively, the storing action depicted in block 52 may be performed in a memory.

[0041] The foregoing actions need not be performed in any particular order (i.e., top to bottom or numerical). For example, actions 40 and 42 may be performed simultaneously or in any order. As another example, design selection 50 may be performed after predicted performance 46 and/or may be performed after generation of design estimates 44 before predicted performance 46.

Additionally, different or fewer actions may be provided. For example, manufacturing action depicted in block 52 need not be performed, such as where the selected design may be used to provide an initial design to be refined by design engineers.

[0042] FIG. 5 illustrates a non-limiting example of machine-learned estimation of sets of design parameters from circuit requirements. Generative machine-learned network 14 may be configured to generate the estimates of the one or more sets of design parameters (e.g., labelled in FIG. 1, as Set 1, Set 2 through Set n) in response to input of the desired circuit requirements (e.g., the desired amplification gain and cutoff frequency of circuit 100 (FIG. 4)). [0043] FIG. 6 illustrates a non-limiting example of machine-learned prediction of respective values of KPIs from the estimated sets of design parameters.

Forward predictive machine-learned network 16 may be configured to generate the predicted values of the one or more key KPIs in response to input of the one or more estimates of sets of design parameters by generative machine-learned network 14.

[0044] FIG. 7 illustrates respective bar plots illustrating four non-limiting examples of possible sets of design parameters estimated (e.g., estimated values of the circuit components) by disclosed embodiments for the circuit illustrated in FIG. 4. That is, FIG. 7 illustrates four different sets of parameter values of circuit components of circuit 100 likely to meet the desired circuit

requirements (e.g., amplification gain and cutoff frequency of circuit 100 (FIG.

4))·

[0045] In one non-limiting embodiment generative machine-learned network may be further configured to estimate respective degrees of influence on the desired circuit requirements due to variation of respective values of parameters involved in respective ones of the one or more sets of estimated design parameters. That is, generative machine-learned network may be configured not just to provide quantitative estimation of the circuit design parameters but also qualitative information (analogous to sensitivity analysis) in connection with the estimated design parameters. Without limitation, bars illustrated in FIG. 7 with a relatively darker (e.g., denser) stippling indicate a relatively larger degree of influence on the desired circuit requirements due to variation of respective values of parameters involved in respective ones of the one or more sets of estimated design parameters. [0046] For example, in Set 2, circuit parameters cl and rl-r3 would have a relatively larger degree of influence on the desired circuit requirements due to variation compared to the other circuit parameters shown in Set 2. By way of comparison, in Set 3, circuit parameters cl and rl would have a relatively larger degree of influence on the desired circuit requirements due to variation compared to the other circuit parameters shown in Set 3.

[0047] FIG. 8 are respective plots illustrating respective performances of the

foregoing sets of estimated design parameters illustrated in FIG. 7 in term of deviations (e.g., errors) of predicted KPIs calculated by disclosed

embodiments relative to simulated true values of the circuit requirements. In one non-limiting embodiment, a relative error may be calculated. It will be appreciated that any user-defined metric or function may be implemented for evaluating the predicted KPIs relative to the desired circuit requirements.

[0048] Without limitation, in certain embodiments the architectures of the respective networks may include convolutional, sub-sampling (e.g., max or average pooling), fully connected layers, recurrent, SoftMax, concatenation, dropout, residual, and/or other types of layers. Any combination of layers may be provided. Any arrangement of layers may be used. Skipping, feedback, or other connections within or between layers may be used. Hierarchical structures may be employed, either for learning features or representation or for classification or regression.

[0049] Without limitation, in certain embodiments the architectures for deep learning may include a convolutional neural network (CNN) or convolution layer. The CNN defines one or more layers where each layer has a filter kernel for convolution with the input to the layer. Other convolutional layers may provide further abstraction, such as receiving the output from a previous layer and convolving the output with another filter kernel. For example, the machine learning may identify filter kernels that extract features conducive to the estimation of the respective sets of design parameters. [0050] Without limitation, in certain embodiments the architecture may include one or more dense layers. The dense layers connect various features from a previous layer to an output from the layer. In certain embodiments, the dense layers may be fully connected layers. One or more fully connected layers may be provided. The dense layers may form a multi-layer perceptron.

[0051] As noted above, in one embodiment, the architecture for the generative model may be a generative adversarial network (GAN) or a mixture density network (mixture of Gaussians or kernel density estimator), or both. The GAN may include an encoder and decoder connected with a bottleneck to generate output information from the input through increasing and then decreasing levels of abstraction. A discriminator may be provided in training to predict the accuracy of the output of the encoder-decoder. This prediction may be fed back to the encoder-decoder in training. Both the discriminator and the encoder-decoder may learn during training. The architecture for the forward predictive model may be a fully-connected residual network. Other types of architectures or corresponding networks may be used, such as other generative models or feedforward neural networks for the generative model or other residual networks for the predictive or simulation model. Other types of machine learning and corresponding networks may be used instead of neural networks and deep learning, such as a support vector machine.

[0052] Without limitation, in certain embodiments the networks may be trained for any number of epochs using any appropriate optimization technique, learning rate, and/or mean-squared error loss. Other arrangements, layers, units, activation functions, architectures, learning rates, optimizations, loss functions, and/or normalization may be used. It will be appreciated that other

architectures may be used in connection with disclosed embodiments. [0053] In operation, disclosed embodiments allow cost-effective and reliable deployment of deep learning algorithms for autonomous electrical circuit design. Without limitation, disclosed embodiments are effective for carrying out substantially automated generative circuit design, such as involving the user to just specify desired properties of the circuit to be designed and evaluation criteria that accept or reject an autonomously designed circuit.

[0054] Disclosed embodiments are believed to be conducive to widespread and

flexible applicability of machine learned networks for generative circuit design. For example, the efficacy of disclosed embodiments is not contingent on the substantial involvement of highly technical personnel with great and specialized expertise in creating complex circuitry. Additionally, disclosed embodiments can make use of machine learned networks that can generate and evaluate a circuit design practically in real time. Lastly, disclosed

embodiments can benefit from parallel processing techniques (such as may involve Graphical processing units (GPUs) configured to perform parallel processing), where the generative design and evaluation may be performed in a relatively constant time regardless of the size of the circuit to be designed.

[0055] While embodiments of the present disclosure have been disclosed in

exemplary forms, it will be apparent to those skilled in the art that many modifications, additions, and deletions can be made therein without departing from the scope of the invention and its equivalents, as set forth in the following claims.

Claims

What is claimed is:

1. A computerized method comprising:

estimating, by way of an artificial intelligence processor (12), one or more sets of design parameters likely to meet desired circuit requirements of a circuit to be designed, the estimating being by a generative machine-learned network (14) in response to input of the desired circuit requirements;

predicting, by way of the artificial intelligence processor (12), values of one or more key performance indicators (KPIs), the predicting being by a forward predictive machine-learned network (16) in response to input of the one or more sets of design parameters estimated by the generative machine- learned network, the predicted values of the one or more KPIs being indicative of respective performance levels of the one or more sets of estimated design parameters relative to the desired circuit requirements;

based on the predicted values of the one or more KPIs of the one or more sets of design parameters estimated by the generative machine-learned network, evaluating whether at least one of the one or more sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements; and

selecting a set of design parameters from the at least one of the one or more sets of estimated design parameters that meets the desired circuit requirements of the circuit to be designed.

2. The computerized method of claim 1, wherein prior to the estimating by the generative machine-learned network, further comprising training the generative machine-learned network with a first dataset obtained from a simulation of a topology of the circuit to be designed, the first dataset configured to train a generative model of the generative machine-learned network to learn estimating the one or more sets of design parameters corresponding to the topology of the circuit to be designed.

3. The computerized method of claim 2, wherein prior to the predicting by the forward predictive machine-learned network, further comprising training the forward predictive machine-learned network with a second dataset obtained from the simulation of the topology of the circuit to be designed, the second dataset configured to train a forward predictive model of the forward predictive machine-learned network to learn predicting the values of the one or more KPIs corresponding to the topology of the circuit to be designed.

4. The computerized method of claim 1, wherein the generative machine- learned network (14) is selected from a group consisting of a mixture density network, a generative adversarial network, both a mixture density network and a generative adversarial network, and a neural network.

5. The computerized method of claim 1, wherein the forward predictive machine-learned network comprises a residual neural network.

6. The computerized method of claim 1, wherein the generative machine-learned network and the forward predictive machine-learned network each comprises a respective feedforward neural network.

7. The computerized method of claim 1, wherein the estimating by the generative machine-learned network (14) further comprises estimating respective degrees of influence on the desired circuit requirements due to variation of respective values of parameters involved in respective ones of the one or more sets of estimated design parameters.

8. The computerized method of claim 1, wherein the evaluating comprises determining respective errors of the predicted values of the one or more KPIs with respect to the desired circuit requirements.

9. The computerized method of claim 8, wherein the determining of respective errors further comprises determining respective relative errors of the predicted values of the one or more KPIs relative to the desired circuit requirements.

10. The computerized method of claim 1, wherein in the event the evaluating determines none of the one or more sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements, triggering the generative machine-learned network to generate further estimates of one or more sets of design parameters likely to meet the desired circuit requirements of the circuit to be designed.

11. The computerized method of claim 10, further triggering the forward predictive machine-learned network to predict further values of the one or more key performance indicators KPIs in response to input of the one or more sets of design parameters further estimated by the generative machine-learned network.

12. A computerized system (10) comprising:

an artificial intelligence processor (12) configured to estimate one or more sets of design parameters likely to meet desired circuit requirements of a circuit to be designed, wherein a generative machine-learned network (14) of the artificial intelligence processor is configured to generate the one or more sets of estimated design parameters in response to input of the desired circuit requirements,

wherein the artificial intelligence processor is further configured to predict values of one or more key performance indicators (KPIs), wherein a forward predictive machine-learned network (16) of the artificial intelligence processor is configured to generate the predicted values of the one or more key KPIs in response to input of the one or more sets of design parameters estimated by the generative machine-learned network, the predicted values of the one or more KPIs being indicative of respective performance levels of the one or more sets of estimated design parameters relative to the desired circuit requirements;

an evaluator (18) configured to evaluate, based on the predicted values of the one or more KPIs, whether at least one of the one or more sets of design parameters estimated by the generative machine-learned network meets the desired circuit requirements,

wherein a set of design parameters is selected from the at least one of the one or more sets of design parameters estimated by the generative machine-learned network that meets the desired circuit requirements of the circuit to be designed.

13. The computerized system of claim 12, wherein prior to generation by the generative machine-learned network of the one or more sets of estimated design parameters, the generative machine-learned network is trained with a first dataset from a simulation of a topology of the circuit to be designed, the first dataset configured to train a generative model of the generative machine- learned network to estimate the one or more sets of design parameters corresponding to the topology of the circuit to be designed.

14. The computerized system of claim 13, wherein prior to generation by the forward predictive machine-learned network of the predicted values of the one or more KPIs, the forward predictive machine-learned network being trained with a second dataset from the simulation of the topology of the circuit to be designed, the second dataset configured to train a forward predictive model of the forward predictive machine-learned network to predict the values of the one or more KPIs corresponding to the topology of the circuit to be designed.

15. The computerized system of claim 12, wherein the generative machine-learned network is selected from the group consisting of a mixture density network, a generative adversarial network, both a mixture density network and a generative adversarial network, and a neural network.

16. The computerized system of claim 12, wherein the forward predictive machine-learned network comprises a residual neural network.

17. The computerized system of claim 12, wherein the generative machine-learned network and the forward predictive machine-learned network each comprises a respective feedforward neural network.

18. The computerized system of claim 12, wherein the generative machine-learned network is further configured to estimate respective degrees of influence on the desired circuit requirements due to variation of respective values of parameters involved in respective ones of the one or more sets of estimated design parameters.

19. The computerized system of claim 12, wherein the evaluator is configured to determine respective relative errors of the predicted values of the one or more KPIs with respect to the desired circuit requirements.

20. The computerized system of claim 12, wherein the circuit to be designed is part of a subsystem comprising a plurality of circuits to be configured with respective sets of design parameters selected from respective sets of design parameters estimated by the generative machine-learned network.