Granted Patent B2

US 12,524,583 · App. 18/706,243 · Granted Jan 13, 2026

Inventors: Suhas Suresha (Menlo Park, CA); Anatoly Aseev (Menlo Park, CA); Alfredo De La Fuente (Menlo Park, CA)

Assignee: Schlumberger Technology Corporation

G06F30/13G06N5/022

Abstract

Claims (63)

1 . A method comprising:

retrieving a plurality of sparse context inputs comprising a plurality of context locations corresponding to a plurality of geological property values for a geological property;

generating, via an attentive neural process model, a predicted mean image for the geological property using a plurality of target locations in a space and the plurality of sparse context inputs, wherein generating the predicted mean image comprises:

encoding, by a latent attention layer of an encoder, the plurality of sparse context inputs to obtain an internal representation, wherein the encoder comprises a deterministic attention layer and the latent attention layer;

parameterizing a distribution of a latent variable using the internal representation;

concatenating the plurality of target locations, an additional internal representation, and the distribution to obtain a concatenated input; and

decoding the concatenated input to generate one or more predicted mean pixels of the predicted mean image and one or more predicted variance pixels of a predicted variance image, wherein each pixel of the predicted mean image comprises a respective target location of the plurality of target locations; and

presenting the predicted mean image.

2 . The method of claim 1 , further comprising:

generating the predicted variance image for the geological property by the attentive neural process model using the plurality of target locations and the plurality of sparse context inputs.

3 . The method of claim 1 , further comprising:

encoding, by the deterministic attention layer of the encoder, the plurality of sparse context inputs to obtain the additional internal representation, the plurality of sparse context inputs comprising the plurality of context locations; and

obtaining weights, by a cross-attention layer, for the plurality of context locations, wherein the weights focus a subset of the plurality of context locations to the plurality of target locations.

4 . The method of claim 1 , further comprising:

processing the predicted mean image and the predicted variance image to generate a predicted sample image.

5 . The method of claim 1 , further comprising:

presenting one or more of the predicted mean image, the predicted variance image, and a predicted sample image generated from the predicted mean image and the predicted variance image.

6 . The method of claim 1 , further comprising:

training the attentive neural process model, wherein training the attentive neural process model comprises:

receiving an original image, wherein the original image comprises a plurality of pixels corresponding to a plurality of geological properties the original image;

generating a plurality of training context locations from the original image, wherein the plurality of training context locations corresponds to a subset of the plurality of pixels of the original image;

processing the original image with the plurality of training context locations;

generating a masked image, wherein the masked image comprises respective geological properties at pixels corresponding to the plurality of training context locations; and

outputting a trained attentive neural process model.

7 . The method of claim 6 , further comprising:

processing the masked image to generate a training mean image and a training variance image;

processing the training mean image and the training variance image to generate a training sample image; and

updating the attentive neural process model based on the training sample image and an original image.

8 . The method of claim 1 , wherein the geological property values are indicative of one or more features of a well comprising an acoustic impedance, a density, a porosity, one or more additional features, or a combination thereof.

9 . The method of claim 1 , further comprising:

generating the predicted variance image, wherein each pixel of the predicted variance image corresponds to a respective target location of the plurality of target locations and a variance value from a mean value for the geological property.

10 . A system comprising:

one or more processors;

memory, accessible by the one or more processors, the memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:

retrieving a plurality of sparse context inputs comprising a plurality of context locations comprising a plurality of geological property values for a geological property;

using an attentive neural process model to process the plurality of sparse context inputs using a plurality of target locations; and

generating, via the attentive neural process model, a predicted mean image for the geological property by the attentive neural process model using the plurality of target locations in a space and the plurality of sparse context inputs, wherein generating the predicted mean image comprises:

encoding, by a latent attention layer of an encoder, the plurality of sparse context inputs to obtain an internal representation, wherein the encoder comprises a deterministic attention layer and the latent attention layer;

parameterizing a distribution of a latent variable using the internal representation;

concatenating the plurality of target locations, an additional internal representation, and the distribution to obtain a concatenated input; and

decoding the concatenated input to generate one or more predicted mean pixels of the predicted mean image and one or more predicted variance pixels of a predicted variance image, and wherein each pixel of the predicted mean image comprises a respective target location of the plurality of target locations.

11 . The system of claim 10 , wherein operations further comprise:

generating the predicted variance image for the geological property by the attentive neural process model using the plurality of target locations and the plurality of sparse context inputs.

12 . The system of claim 10 , wherein operations further comprise:

encoding, by the deterministic attention layer of the encoder, the plurality of sparse context inputs to obtain the additional internal representation, the plurality of sparse context inputs comprising the plurality of context locations; and

obtaining weights, by a cross-attention layer, for the plurality of context locations, wherein the weights focus a subset of the plurality of context locations to the plurality of target locations.

13 . The system of claim 10 , wherein operations further comprise:

processing the predicted mean image and the predicted variance image to generate a predicted sample image.

14 . The system of claim 10 , wherein operations further comprise:

presenting one or more of the predicted mean image, the predicted variance image and a predicted sample image generated from the predicted mean image and the predicted variance image.

15 . The system of claim 10 , wherein operations further comprise:

processing with the attentive neural process model, wherein the attentive neural process model is trained by:

receiving an original image;

selecting a plurality of training context locations from the original image; and

processing the original image with the training context locations to generate a masked image.

16 . A non-transitory computer readable medium comprising instructions that, when executed by one or more processors, are configured for:

retrieving a plurality of sparse context inputs comprising a plurality of context locations corresponding to a plurality of geological property values for a geological property;

generating, via an attentive neural process model, a predicted mean image for the geological property using a plurality of target locations and the plurality of sparse context inputs, wherein generating the predicted mean image comprises:

encoding, by a latent attention layer of an encoder, the plurality of sparse context inputs to obtain an internal representation, wherein the encoder comprises a deterministic attention layer and the latent attention layer;

parameterizing a distribution of a latent variable using the internal representation;

concatenating the plurality of target locations, an additional internal representation, and the distribution to obtain a concatenated input; and

decoding the concatenated input to generate one or more predicted mean pixels of the predicted mean image and one or more predicted variance pixels of a predicted variance image, wherein each pixel of the predicted mean image comprises a respective target location of the plurality of target locations; and

presenting the predicted mean image.

Assignments (1)

Continuity (2)

References Cited (43)