Superresolution and consistency constraints to scale up deep learning models

Patent metadata

Field	Value
Publication number	US-12169776-B2
Application number	US-202017121933-A
Country	US
Kind code	B2
Filing date	Dec 15, 2020
Priority date	Dec 15, 2020
Publication date	Dec 17, 2024
Grant date	Dec 17, 2024

Abstract

Techniques of facilitating deep learning model rescaling by computing devices. In one example, a system can comprise a processor that executes computer executable components stored in memory. The computer executable components can comprise: a rescaling component; and a forecasting component. The rescaling component can determine a scaling ratio that maps low mesh resolution predictive data output by a partial differential equation (PDE)-based model for a sub-domain to high-resolution observational or ground-truth data for a domain comprising the sub-domain. The forecasting component can generate high mesh resolution predictive data for the domain with a machine-learning model using input data of the PDE-based model and the scaling ratio.

First claim

What is claimed is:

1. A system, comprising:
at least one processor; and
at least one memory that stores executable instructions that, when executed by the at least one processor, facilitate performance of operations, comprising:
determining a scaling ratio that maps low mesh resolution predictive data output by a partial differential equation (PDE)-based model for a sub-domain to high-resolution observational or ground-truth data for a domain comprising the sub-domain,
wherein determining the scaling ratio comprises:
extracting patches, comprising overlapping boundaries, from the low mesh resolution predictive data and representing each extracting patch as a high-dimensional vector,
non-linearly mapping each high-dimensional vector onto another high-dimensional vector from the high-resolution observational or ground-truth data, and
aggregating high-resolution patch-wise representations corresponding to each non-linearly mapped vector to generate the high-resolution observational or ground-truth data; and

generating high mesh resolution predictive data for the domain with a machine-learning model using input data of the PDE-based model and the scaling ratio.

2. The system of claim 1 , wherein the operations further comprise:
determining a plurality of scaling ratios that map predictive data output by different PDE-based models for different sub-domains at different mesh resolutions to the high-resolution observational or ground-truth data.

3. The system of claim 1 , wherein the operations further comprise:
generating the high mesh resolution predictive data for the domain with the machine-learning model using input data for a plurality of PDE-based models at different mesh resolutions for different sub-domains comprising the domain.

4. The system of claim 1 , wherein the operations further comprise:
generating consistency constraints to enforce neighboring synchronization at interfaces between different sub-domains comprising the domain.

5. The system of claim 4 , wherein the consistency constraints include: consistency constraints from a high mesh resolution PDE-based model; consistency constraints from a low mesh resolution PDE-based model; consistency constraints from adjacent tiles of a common mesh resolution PDE-based model; or a combination thereof.

6. The system of claim 4 , wherein the consistency constraints define bounds on high mesh resolution predictive data values output by the machine-learning model at selected points, bounds on a modulus of continuity, bounds on sub-gradients, bounds on a sum of sub-gradients across a tile of predictive data, or a combination thereof.

7. The system of claim 4 , wherein the operations further comprise:
generating the consistency constraints using sensor data corresponding to the domain, a total variance across a patch of a mesh, or a combination thereof.

8. The system of claim 1 , wherein the operations further comprise:
using machine learning to train the machine-learning model using a data set comprising historical input-output pairs of the PDE-based model.

9. The system of claim 8 , wherein the data set further comprises additional inputs generated by the machine-learning model.

10. A computer-implemented method, comprising:
determining, by a system operatively coupled to a processor, a scaling ratio that maps low mesh resolution predictive data output by a partial differential equation (PDE)-based model for a sub-domain to high-resolution observational or ground-truth data for a domain comprising the sub-domain,
wherein determining the scaling ratio comprises:
extracting patches, comprising overlapping boundaries, from the low mesh resolution predictive data and representing each extracting patch as a high-dimensional vector,
non-linearly mapping each high-dimensional vector onto another high-dimensional vector from the high-resolution observational or ground-truth data, and
aggregating high-resolution patch-wise representations corresponding to each non-linearly mapped vector to generate the high-resolution observational or ground-truth data; and

generating, by the system, high mesh resolution predictive data for the domain with a machine-learning model using input data of the PDE-based model and the scaling ratio.

11. The computer-implemented method of claim 10 , wherein the system determines a plurality of scaling ratios that map predictive data output by different PDE-based models for different sub-domains at different mesh resolutions to the high-resolution observational or ground-truth data.

12. The computer-implemented method of claim 10 , wherein the system generates the high mesh resolution predictive data for the domain with the machine-learning model using input data for a plurality of PDE-based models at different mesh resolutions for different sub-domains comprising the domain.

13. The computer-implemented method of claim 10 , further comprising:
generating, by the system, consistency constraints to enforce neighboring synchronization at interfaces between different sub-domains comprising the domain.

14. The computer-implemented method of claim 13 , wherein the consistency constraints include: consistency constraints from a high mesh resolution PDE-based model; consistency constraints from a low mesh resolution PDE-based model; consistency constraints from adjacent tiles of a common mesh resolution PDE-based model; or a combination thereof.

15. The computer-implemented method of claim 13 , wherein the consistency constraints define bounds on high mesh resolution predictive data values output by the machine-learning model at selected points, bounds on a modulus of continuity, bounds on sub-gradients, bounds on a sum of sub-gradients across a tile of predictive data, or a combination thereof.

16. The computer-implemented method of claim 13 , wherein the system generates the consistency constraints using sensor data corresponding to the domain, a total variance across a patch of a mesh, or a combination thereof.

17. The computer-implemented method of claim 10 , further comprising:
employing, by the system, machine learning to train the machine-learning model using a data set comprising historical input-output pairs of the PDE-based model, additional inputs generated by the machine-learning model, or a combination thereof.

18. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
determine, by the processor, a scaling ratio that maps low mesh resolution predictive data output by a partial differential equation (PDE)-based model for a sub-domain to high-resolution observational or ground-truth data for a domain comprising the sub-domain,
wherein determining the scaling ratio comprises:
extracting patches, comprising overlapping boundaries, from the low mesh resolution predictive data and representing each extracting patch as a high-dimensional vector,
non-linearly mapping each high-dimensional vector onto another high-dimensional vector from the high-resolution observational or ground-truth data, and
aggregating high-resolution patch-wise representations corresponding to each non-linearly mapped vector to generate the high-resolution observational or ground-truth data; and

generate, by the processor, high resolution predictive data for the domain with a machine-learning model using input data of the PDE-based model and the scaling ratio.

19. The computer program product of claim 18 , the program ins

Assignees

• IBM

Inventors

• O'Donncha Fearghal
• RAWAT Ambrish
• Mckenna Sean A
• Sinn Mathieu

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06T3/4053

  based on super-resolution, i.e. the output image resolution being higher than the sensor resolution · CPC title