	US 11,714,937 B2
	Estimating physical parameters of a physical system based on a spatial-temporal emulator
Adrian Albert, Berkeley, CA (US)
Assigned to Terrafuse, Inc., San Francisco, CA (US)
Filed by Terrafuse, Inc., Berkeley, CA (US)
Filed on Nov. 15, 2021, as Appl. No. 17/526,057.
Application 17/526,057 is a continuation of application No. 16/550,234, filed on Aug. 25, 2019, granted, now 11,205,028.
Claims priority of provisional application 62/727,992, filed on Sep. 6, 2018.
Claims priority of provisional application 62/728,000, filed on Sep. 6, 2018.
Prior Publication US 2022/0075913 A1, Mar. 10, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 30/20 (2020.01); G06N 20/00 (2019.01); G06F 111/10 (2020.01)

CPC G06F 30/20 (2020.01) [G06N 20/00 (2019.01); G06F 2111/10 (2020.01)]

19 Claims

1. A method for generating simulations of physical variables of a physical system, comprising:

obtaining observational data, wherein the observational data includes at least one source of physical data, comprising one or more sensors sensing the physical data;

obtaining, by one or more computing devices or storage devices that are connected through one or more networks, numeric simulation data;

fusing, by one or more computing devices, the observation data and the numeric simulation data, comprising:

preprocessing the observational data and the numeric simulation data to remove inconsistencies of the observational data and the numeric simulation data;

processing the preprocessed observational data and the numeric simulation data to extract interpretable structures and patterns within that data using ground truth and labeled information to create domain interpretable data;

normalizing the preprocessed observation data, the numeric simulation data, and the domain interpretable data layers;

increasing a resolution of a gridding of the normalized preprocessed observation data, numeric simulation data, and domain interpretable data layers;

training a spatial-temporal emulator model for a physical numerical model using the normalized preprocessed observation data, the numeric simulation data, and the domain interpretable data;

incorporating prior knowledge of the physical system into the spatial-temporal emulator model;

the method further comprising estimating, by one or more computing devices, one or more physical parameters of the physical system based on the trained spatial-temporal emulator model;

compressing the trained spatial-temporal emulator model, comprising:

1) generating candidate mutations of an architecture of the trained spatial-temporal emulator model to reduce a number of parameters or connections of the parameters;

2) evaluating a performance of each of the candidate mutations of the architecture on validation data using metrics;

3) retaining a subset of candidate mutations exhibiting best performance on the metrics; and

4) iterating steps 1-3 until convergence to a desired reduction in size of the trained spatial-temporal emulator model, yielding a compressed trained spatial-temporal model; and

the method further comprising utilizing, by one or more computing devices, the estimated one or more physical parameters for at least one of a plurality of applications.