Capacitance extraction systems based on machine learning models
Extraction of capacitance values from a design of an electrical circuit can use a set of trained neural networks to generate extracted capacitance values from the circuit using a representation of the Green's function. A method can include the following operations: storing a machine learning model that includes a trained set of one or more neural networks that have been trained to calculate a representation of a Green's function to extract capacitance values from a design of a circuit having a set of conductors; applying, to the machine learning model, a set of inputs representing the set of conductors and their surrounding dielectric materials; encoding the set of inputs through a trained encoder to generate a latent space representation; calculating the values of the Green's function from the latent space representation through a dedicated trained neural network; and calculating the values of the gradient of the Green's function from the latent space representation through another dedicated trained neural network.
1. A non-transitory machine readable medium storing executable instructions that when executed by a data processing system cause the data processing system to perform a method comprising:
storing a machine learning model that includes a trained set of one or more neural networks that have been trained to represent calculations for a Green's function;
applying, to the machine learning model, a set of inputs representing a set of dielectric materials present within volumes inside a first design of a circuit;
encoding the set of inputs through a trained encoder in the one or more neural networks to generate a latent space representation as an encoding of permittivity values of the set of inputs;
calculating the Green's function of the set of inputs from the latent space representation through one of the one or more neural networks trained for calculation of the Green's function; and
calculating a gradient of the Green's function of the set of inputs from the latent space representation through the one or more neural networks trained for calculation of the gradient of the Green's function.
2. The non-transitory machine readable medium as in claim 1 , wherein a representation of the Green's function is generated, at inference time, during a random walk among a set of conductors and wherein values of the gradient of the Green's function are calculated during the random walk.
3. The non-transitory machine readable medium as in claim 2 , wherein the trained encoder encodes characteristics of dielectric permittivity values present within volumes inside the first design of the circuit, and wherein the representation in the latent space is equivariant to rotation of these volumes.
4. The non-transitory machine readable medium as in claim 3 , wherein the one or more neural networks trained for calculation of the gradient of the Green's function comprises a neural network trained for a first axis that is also used for gradient computations in a second axis and a third axis after transformations for the second axis and the third axis in order to reduce a size of the machine learning model.
5. The non-transitory machine readable medium as in claim 4 , wherein the trained set of one or more neural networks comprises a first face decoder that generates a first Green's function output for a first face of a volume, and the first face decoder is used with a set of transformations of the representation in the latent space to generate a set of Green's function outputs for a second face, a third face, a fourth face, a fifth face, and a sixth face of the volume.
6. The non-transitory machine readable medium as in claim 3 , wherein the representation of the Green's function is a Gaussian mixture model that approximates the Green's function for a selected face of a transition volume.
7. The non-transitory machine readable medium as in claim 6 , wherein a trained mixture density neural network in the trained set of one or more neural networks infers parameters of the Gaussian mixture model to approximate the Green's function on the selected face.
8. The non-transitory machine readable medium as in claim 7 , wherein the inferred parameters of the Gaussian mixture model are used to pick random samples from an approximation of the Green's function for the selected face.
9. The non-transitory machine readable medium as in claim 8 , wherein a trained face selector neural network, in the one or more neural networks, receives the representation in the latent space and generates a face distribution from which a face is selected, and the representation in the latent space is transformed based on the selected face and applied as an input to the trained mixture density neural network.
10. The non-transitory machine readable medium as in claim 9 , wherein the gradient of the Green's function is calculated only once over a path of the random walk.
11. A non-transitory machine readable medium storing executable instructions that when executed by a data processing system cause the data processing system to perform a method comprising:
training an encoder to generate encoded outputs, in a latent space, from a set of inputs representing sets of dielectric materials present within volumes inside a design of a circuit;
training a set of neural networks to calculate a representation of Green's function to extract capacitance values from conductors in a set of one or more training circuits; and
storing a machine learning model that includes the trained encoder and the trained set of neural networks.
12. The non-transitory machine readable medium as in claim 11 , wherein Green's function values are calculated, at inference time, during a random walk among a set of conductors in a circuit and wherein values of a gradient of the Green's function values are calculated during the random walk.
13. The non-transitory machine readable medium as in claim 12 , wherein the trained encoder encodes characteristics of dielectric permittivity values present within volumes inside the design of the circuit, and wherein a latent space representation from the trained encoder is equivariant to rotation of these volumes.
14. The non-transitory machine readable medium as in claim 13 , wherein the neural network for calculation of the gradient of the Green's function values includes a neural network trained for a first axis that is also used for gradient computations in a second axis and a third axis after transformations for the second axis and the third axis in order to reduce a size of the machine learning model.
15. The non-transitory machine readable medium as in claim 14 , wherein the trained set of neural networks comprises a first face decoder that generates a first Green's function output for a first face of a volume, and the first face decoder is used with a set of transformations of the latent space representation to generate a set of Green's function outputs for a second face, a third face, a fourth face, a fifth face, and a sixth face of the volume.
16. The non-transitory machine readable medium as in claim 13 , wherein the representation of Green's function is a Gaussian mixture model that approximates the Green's function for a selected face of a transition volume.
17. The non-transitory machine readable medium as in claim 16 , wherein a trained mixture density neural network in the trained set of neural networks infers parameters of the Gaussian mixture model to approximate the Green's function on the selected face.
18. The non-transitory machine readable medium as in claim 17 , wherein the inferred parameters of the Gaussian mixture model are used to pick random samples from an approximation of the Green's function for the selected face.
19. The non-transitory machine readable medium as in claim 18 , wherein a trained face selector neural network receives the latent space representation and generates a face distribution from which a face is selected, and the latent space representation is transformed based on the selected face and applied as an input to the trained mixture density neural network.
20. The non-transitory machine readable medium as in claim 19 , wherein a gradient output is calculated only once over a path of the random walk.
21. The non-transitory machine readable medium as in claim 13 , wherein the encoder is trained as an autoencoder and is trained separately from the set of neural networks that calculate a representation of Green's function.
22. A system comprising:
a memory storing instructions; and
one or more processors coupled to the memory, the one or more processors executing the instructions from the memory to perform a method comprising:
storing a machine learning model that includes a trained set of one or more neural networks that have been trained for calculations of a Green's function;
applying, to the machine learning model, a set of inputs representing a set of dielectric materials present within volumes inside a design of a circuit;
encoding the set of inputs through a trained encoder in the one or more neural networks to generate a latent space representation as an encoding of permittivity values of the set of inputs;
calculating the Green's function of the set of inputs from the latent space representation through one of the one or more neural networks; and
calculating a gradient of the Green's function based on the set of inputs from the latent space representation through the one or more neural networks.