Granted Patent B2

US 12,299,364 · App. 17/245,306 · Granted May 13, 2025

Inventors: Nikhil Gupta (Plano, TX); Timothy W. Fischer (McKinney, TX); Ashish Khandelwal (Frisco, TX); Sreenivasan K. Koduri (Dallas, TX)

Assignee: TEXAS INSTRUMENTS INCORPORATED

G06F30/27G06F18/24155G06F30/3308G06F30/337G06F30/367G06F30/373G06F30/392G06F30/398G06N3/045G06N3/08G06N3/084G06N20/00G06N20/20G06F2111/04G06F2111/20

Abstract

Claims (78)

1. A method comprising:

producing a first machine learning (ML) model, including receiving an initial set of parameters, wherein the initial set of parameters represents inputs to the first ML model, the initial set of parameters associated with a sub-circuit;

interacting a first parameter of the initial set of parameters with other parameters of the initial set of parameters to generate a set of interacted parameters;

adding the interacted parameter to the initial set of parameters to generate a candidate set of parameters;

performing a linear regression on parameters of the candidate set of parameters against a set of expected parameter values to determine a predictive value for parameters of the candidate set of parameters;

removing parameters of the candidate set of parameters based on a comparison between the predictive value and a predetermined predictive threshold;

determining an accuracy of the candidate set of parameters based on the set of expected parameter values;

comparing the accuracy of the candidate set of parameters to a predetermined accuracy level;

determining that the accuracy of the candidate set of parameters reaches the predetermined accuracy level;

in response to determining that the accuracy of the candidate set of parameters reaches the predetermined accuracy level, storing the candidate set of parameters as a set of inputs for a first layer of the first ML model for the sub-circuit for a process technology; and

training the first ML model using the candidate set of parameters stored as the set of inputs, thereby producing a trained ML model.

2. The method of claim 1 , wherein determining the predictive value for the parameters of the candidate set of parameters includes determining a statistical significance value for each parameter of the candidate set of parameters, and wherein removing the parameters is based on a comparison between the statistical significance value for each parameter of the candidate set of parameters and the predetermined predictive threshold.

3. The method of claim 1 , wherein the initial set of parameter includes an interaction parameter modeling how two or more other parameters, of the initial set of parameters, interact.

4. The method of claim 3 , wherein the interaction parameter is based on a circuit theory equation.

5. The method of claim 1 , wherein the initial set of parameters includes one or more sub-circuit physical parameters associated with the sub-circuit and wherein the set of expected parameter values includes one or more sub-circuit performance parameters associated with the sub-circuit.

6. The method of claim 1 , wherein the initial set of parameters include one or more sub-circuit performance parameters associated with the sub-circuit and wherein the set of expected parameter values includes one or more sub-circuit physical parameters associated with the sub-circuit.

7. The method of claim 1 , wherein the initial set of parameters includes one or more parameter values based on properties of the process technology.

8. A non-transitory program storage device comprising instructions stored thereon to cause one or more processors to:

produce a first machine learning (ML) model, including receiving an initial set of parameters, wherein the initial set of parameters represents inputs to the first ML model, the initial set of parameters associated with a sub-circuit;

interact a first parameter of the initial set of parameters with other parameters of the initial set of parameters to generate a set of interacted parameters;

add the interacted parameter to the initial set of parameters to generate a candidate set of parameters;

perform a linear regression on parameters of the candidate set of parameters against a set of expected parameter values to determine a predictive value for parameters of the candidate set of parameters;

remove parameters of the candidate set of parameters based on a comparison between the predictive value and a predetermined predictive threshold;

determine an accuracy of the candidate set of parameters based on the set of expected parameter values;

compare the accuracy of the candidate set of parameters to a predetermined accuracy level;

wherein if the accuracy of the candidate set of parameters reaches the predetermined accuracy level, store the candidate set of parameters as a first set of inputs for a first layer of the first ML model for the sub-circuit for a process technology; and

wherein if the accuracy of the candidate set of parameters does not reach the predetermined accuracy level, perform at least one more iteration of interaction to add to the candidate set of parameters and linear regression to remove from the candidate set of parameters until the accuracy of the candidate set of parameters has reached the predetermined accuracy level and then store the candidate set of parameters as a second set of inputs for the first layer of the first ML model for the sub-circuit for the process technology; and

train the first ML model using the candidate set of parameters stored as the first or second set of inputs, thereby producing a trained ML model.

9. The non-transitory program storage device of claim 8 , wherein the predictive value for the parameters of the candidate set of parameters includes a statistical significance value for each parameter of the candidate set of parameters, and wherein the instructions to cause the one or more processors to remove parameters of the candidate set of parameters includes instructions to cause the one or more processors to remove parameters based on a comparison between the statistical significance value for each parameter of the candidate set of parameters and the predetermined predictive threshold.

10. The non-transitory program storage device of claim 8 , wherein the initial set of parameter includes an interaction parameter modeling how two or more other parameters, of the initial set of parameters, interact.

11. The non-transitory program storage device of claim 10 , wherein the interaction parameter is based on a circuit theory equation.

12. The non-transitory program storage device of claim 8 , wherein the initial set of parameters includes one or more sub-circuit physical parameters associated with the sub-circuit and wherein the set of expected parameter values includes one or more sub-circuit performance parameters associated with the sub-circuit.

13. The non-transitory program storage device of claim 8 , wherein the initial set of parameters include one or more sub-circuit performance parameters associated with the sub-circuit and wherein the set of expected parameter values includes one or more sub-circuit physical parameters associated with the sub-circuit.

14. The non-transitory program storage device of claim 8 , wherein the initial set of parameters includes one or more parameter values based on properties of the process technology.

15. An electronic device, comprising:

a memory; and

one or more processors operatively coupled to the memory, wherein the one or more processors are configured to execute instructions causing the one or more processors to:

produce a first machine learning (ML) model, including receiving an initial set of parameters, wherein the initial set of parameters represents inputs to the first ML model, the initial set of parameters associated with a sub-circuit;

interact a first parameter of the initial set of parameters with other parameters of the initial set of parameters to generate a set of interacted parameters;

add the interacted parameter to the initial set of parameters to generate a candidate set of parameters;

perform a linear regression on parameters of the candidate set of parameters against a set of expected parameter values to determine a predictive value for parameters of the candidate set of parameters;

remove parameters of the candidate set of parameters based on a comparison between the predictive value and a predetermined predictive threshold;

determine an accuracy of the candidate set of parameters based on the set of expected parameter values;

compare the accuracy of the candidate set of parameters to a predetermined accuracy level;

wherein if the accuracy of the candidate set of parameters reaches the predetermined accuracy level, store the candidate set of parameters as a first set of inputs for a first layer of the first ML model for the sub-circuit for a process technology; and

wherein if the accuracy of the candidate set of parameters does not reach the predetermined accuracy level, perform at least one more iteration of interaction to add to the candidate set of parameters and linear regression to remove from the candidate set of parameters until the accuracy of the candidate set of parameters has reached the predetermined accuracy level and then store the candidate set of parameters as a second set of inputs for the first layer of the first ML model for the sub-circuit for the process technology; and

train the first ML model using the candidate set of parameters stored as the first or second set of inputs, thereby producing a trained ML model.

16. The electronic device of claim 15 , wherein the predictive value for the parameters of the candidate set of parameters includes a statistical significance value for each parameter of the candidate set of parameters, and wherein the instructions to cause the one or more processors to remove parameters of the candidate set of parameters includes instructions to cause the one or more processors to remove parameters based on a comparison between the statistical significance value for each parameter of the candidate set of parameters and the predetermined predictive threshold.

17. A method comprising:

producing a first machine learning (ML) model, including receiving an initial set of parameters, wherein the initial set of parameters represents inputs to the first ML model, the initial set of parameters associated with a sub-circuit;

interacting a first parameter of the initial set of parameters with other parameters of the initial set of parameters to generate a set of interacted parameters;

adding the interacted parameter to the initial set of parameters to generate a candidate set of parameters;

performing a linear regression on parameters of the candidate set of parameters against a set of expected parameter values to determine a predictive value for parameters of the candidate set of parameters;

removing parameters of the candidate set of parameters based on a comparison between the predictive value and a predetermined predictive threshold;

determining an accuracy of the candidate set of parameters based on the set of expected parameter values;

comparing the accuracy of the candidate set of parameters to a predetermined accuracy level;

determining that the accuracy of the candidate set of parameters does not reach the predetermined accuracy level;

in response to determining that the accuracy of the candidate set of parameters does not reach the predetermined accuracy level, performing at least one more iteration of interaction to add to the candidate set of parameters and linear regression to remove from the candidate set of parameters until the accuracy of the candidate set of parameters has reached the predetermined accuracy level;

storing the candidate set of parameters as a set of inputs for a first layer of the first ML model for the sub-circuit for a process technology; and

training the first ML model using the candidate set of parameters stored as the set of inputs, thereby producing a trained ML model.

18. The method of claim 17 , wherein determining the predictive value for the parameters of the candidate set of parameters includes determining a statistical significance value for each parameter of the candidate set of parameters, and wherein removing the parameters is based on a comparison between the statistical significance value for each parameter of the candidate set of parameters and the predetermined predictive threshold.

19. A method comprising:

producing a first machine learning (ML) model, including receiving an initial set of parameters, wherein the initial set of parameters represents inputs to the first ML model, the initial set of parameters associated with a sub-circuit;

interacting a first parameter of the initial set of parameters with other parameters of the initial set of parameters to generate a set of interacted parameters;

adding the interacted parameter to the initial set of parameters to generate a candidate set of parameters;

performing a linear regression on parameters of the candidate set of parameters against a set of expected parameter values to determine a predictive value for parameters of the candidate set of parameters;

removing parameters of the candidate set of parameters based on a comparison between the predictive value and a predetermined predictive threshold;

determining an accuracy of the candidate set of parameters based on the set of expected parameter values;

comparing the accuracy of the candidate set of parameters to a predetermined accuracy level;

determining that the accuracy of the candidate set of parameters does not reach the predetermined accuracy level;

in response to determining that the accuracy of the candidate set of parameters does not reach the predetermined accuracy level, performing at least one more iteration of interaction to add to the candidate set of parameters and linear regression to remove from the candidate set of parameters until each parameter of the initial set of parameters has been interacted with other parameters of the candidate set a predetermined number of times;

training a second ML model based on the candidate set of parameters and the set of expected parameter values;

tuning a hyperparameter of the second ML model;

determining the accuracy for the second ML model;

determining that the accuracy of the second ML is greater than the predetermined accuracy level; and

storing the candidate set of parameters as a set of inputs for the second ML model for the sub-circuit for a process technology; and

training the second ML model using the candidate set of parameters stored as the set of inputs, thereby producing a trained ML model.

20. The method of claim 19 , wherein tuning the hyperparameter comprises applying Bayesian hyperparameter optimization to the hyperparameter.

Assignments (2)

Continuity (3)

References Cited (45)