US 12,423,610 B2
Estimation of global thermal conditions via cosimulation of machine learning outputs and observed data
Zainab Diana Titus, London (GB); Annabel Causer, Exeter (GB); Graham Baines, Abingdon (GB); Christine Yallup, Abingdon (GB); Olutobi Adeyemi, Didcot (GB); and Marc Paul Servais, Reading (GB)
Assigned to Landmark Graphics Corporation, Houston, TX (US)
Filed by Landmark Graphics Corporation, Houston, TX (US)
Filed on Jul. 15, 2020, as Appl. No. 16/929,191.
Prior Publication US 2022/0019933 A1, Jan. 20, 2022
Int. Cl. G06N 20/00 (2019.01); G01K 17/08 (2006.01); G06F 18/15 (2023.01); G06F 18/214 (2023.01)
CPC G06N 20/00 (2019.01) [G01K 17/08 (2013.01); G06F 18/15 (2023.01); G06F 18/214 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
receiving a plurality of aggregated observed heat flow values and geological data for an earth formation;
processing the observed heat flow values and geological data to filter out unreliable data using statistical measures, wherein unreliable data includes data points more than a specified statistical measure away from a mean value of the data;
normalizing the processed observed heat flow values and geological data, wherein normalizing the data includes mapping categorical geological data onto numerical data using an integer scale with a number of entries identical to a number of categories for each categorical geological parameter;
training each supervised learning model in a plurality of supervised learning models to estimate a plurality of heat flow values, wherein the training includes,
separating the normalized observed heat flow values and geological data into training data and test data,
generating estimated heat flow values from the test data,
comparing the estimated heat flow values with the normalized data, and
propagating the value of a loss function of the estimated heat flow values and the normalized heat flow values backward through layers of each model using gradient descent to update the model based on the comparison of the estimated heat flow values with the normalized data;
evaluating errors of each of the trained supervised learning models using the training data and the test data;
selecting a first selected model meeting according to a corresponding model selection criterion, wherein the model selection criteria includes at least selecting a model having a lowest weighted average of errors;
cosimulating the estimated plurality of heat flow values of the first selected model and the plurality of observed heat flow values to generate a cosimulated heat flow map; and
using the cosimulated heat flow map as input to one or more processing operations for facilitating subsurface formation evaluation.