Compressed field response representation for memory efficient physical device simulation

Patent metadata

Field	Value
Publication number	US-11501169-B1
Application number	US-201916398697-A
Country	US
Kind code	B1
Filing date	Apr 30, 2019
Priority date	Apr 30, 2019
Publication date	Nov 15, 2022
Grant date	Nov 15, 2022

Abstract

A method of optimizing structural parameters of a physical device includes: receiving an initial description of the physical device that describes the physical device with an array of voxels that each describe one or more of the structural parameters; performing a time-forward simulation of a field response propagating through the physical device and interacting with the voxels in a simulated environment, wherein the field response is influenced by the structural parameters of the voxels; generating field response values describing the field response at each of the voxels for each of a plurality of time steps; encoding the field response values to generate compressed field response values; storing the compressed field response values; decoding one or more of the compressed field response values to extract regenerated field response values; and generating a revised description of the physical device having a structural parameter optimized.

First claim

What is claimed is:

1. A computer-implemented method of optimizing structural parameters of a physical device, the method comprising:
receiving an initial description of the physical device that describes the physical device with an array of voxels that each describe one or more of the structural parameters of the physical device in a corresponding location in an N dimensional space, wherein N=1, 2, 3, or more;
performing a time-forward simulation of a field response propagating through the physical device and interacting with the voxels in a simulated environment, wherein the field response is generated in response to an excitation source in the simulated environment, wherein the field response is influenced by the structural parameters of the voxels;
generating field response values describing the field response at each of the voxels for each of a plurality of time steps of the time-forward simulation;
encoding the field response values to generate compressed field response values;
storing the compressed field response values with a reduced memory footprint relative to storing the field response values;
decoding one or more of the compressed field response values to extract regenerated field response values; and
generating a revised description of the physical device having at least one of the structural parameters of at least one of the voxels optimized based at least upon the regenerated field response values.

2. The computer-implemented method of claim 1 , wherein generating the revised description of the physical device comprises performing a backpropagation structural optimization using the regenerated field response values, the backpropagation structural optimization comprising:
computing a loss value based at least on a difference between the field response and a target field response;
performing a backwards simulation of a reverse field response that back propagates the loss value through the voxels in the simulated environment;
computing a field gradient for a given voxel based upon one or more of the regenerated field response values;
computing a loss gradient for the given voxel based upon the reverse field response;
generating a structural gradient for the given voxel by combining the field gradient with the loss gradient; and
revising one or more of the structural parameters at the given voxel based at least upon the structural gradient.

3. The computer-implemented method of claim 2 , wherein the loss value is based upon the difference between the field response and the target field response at a designated time step and a designated location within the simulated environment.

4. The computer-implemented method of claim 1 , wherein encoding the field response values to generate the compressed field response values comprises:
encoding the field response values with a learned autoencoder.

5. The computer-implemented method of claim 4 , wherein encoding the field response values to generate the compressed field response values further comprises:
inputting structural features of the physical device into the learned autoencoder to leverage constraints that the structural features place on the field response to further reduce the memory footprint associated with the compressed field response values.

6. The computer-implemented method of claim 4 , further comprising:
initializing weights associated with periodic activation functions of the learned autoencoder to correspond to expected spatial frequencies in the field response, wherein the expected spatial frequencies are determined based at least upon physical dimensions of structural features in the physical device that influence the field response.

7. The computer-implemented method of claim 1 , wherein encoding the field response values to generate the compressed field response values comprises:
encoding the field response values with a spatial Fourier transform based compression scheme.

8. The computer-implemented method of claim 7 , wherein the spatial Fourier transform based compression scheme comprises a JPEG compressing scheme.

9. The computer-implemented method of claim 1 , wherein encoding the field response values to generate compressed field response values comprises:
encoding the field response values with a motion-based compression scheme that uses keyframes at determined time steps and delta frames between the keyframes describing changes in the field response values relative to the keyframes.

10. The computer-implemented method of claim 9 , further comprising:
executing an autoencoder to evaluate when to create a given one of the keyframes at one of the determined time steps after a consecutive set of the delta frames,
wherein a number of the delta frames between consecutive ones of the keyframes is variable based upon evaluation of the autoencoder.

11. The computer-implemented method of claim 1 , wherein encoding the field response values to generate compressed field response values comprises:
spatially sub-sampling the field response values in a first region of the simulated environment where the field response values have a lower spatial variation than another region of the simulated environment.

12. At least one non-transitory computer-readable storage medium that provides instructions that, when executed by a machine, will cause the machine to perform operations comprising:
receiving an initial description of a physical device that describes the physical device with an array of voxels that each describe one or more structural parameters of the physical device in a corresponding location in a N-dimensional space, wherein N=1, 2, 3, or more;
performing a time-forward simulation of a field response propagating through the physical device and interacting with the voxels in a simulated environment, wherein the field response is generated in response to an excitation source in the simulated environment, wherein the field response is influenced by the structural parameters of the voxels;
generating field response values describing the field response at each of the voxels for each of a plurality of time steps of the time-forward simulation;
encoding the field response values to generate compressed field response values;
storing the compressed field response values with a reduced memory footprint relative to storing the field response values;
decoding one or more of the compressed field response values to extract regenerated field response values; and
generating a revised description of the physical device having at least one of the structural parameters of at least one of the voxels optimized based at least upon the regenerated field response values.

13. The least one non-transitory computer-readable storage medium of claim 12 , wherein generating the revised description of the physical device comprises performing a backpropagation structural optimization, the backpropagation structural optimization comprising:
computing a loss value based at least on a difference between the field response and a target field response;
performing a backwards simulation of a reverse field response that back propagates the loss value through the voxels in the simulated environment;
computing a field gradient for a given voxel based upon one or more of the regenerated field response values;
computing a loss gradient for the given voxel based upon the reverse field response;
generating a structural gradient for the given voxel by combining the field gradient with the loss gradient; and
revising one or more of the structural parameters at the given voxel based at least upon the structural gradient.

14. The least one non-transitory computer-rea

Assignees

• X Dev Llc

Inventors

• Schubert Martin
• Adolf Brian
• Lu Jesse

Classifications

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/0455 Primary

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06F30/27

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title

• G06F30/20

  Design optimisation, verification or simulation (optimisation, verification or simulation of circuit designs G06F30/30) · CPC title

• G06T12/30

  Image post-processing, e.g. metal artefact correction · CPC title