Machine Learning Model Selection for Lithology Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inaccurate classifications of subterranean formations due to differing probability distributions between training and test data sets lead to poor hydrocarbon production, as existing methods fail to effectively select suitable machine-learning models for lithology classification.
Innovation Solution
A system that calculates similarity between training and test data sets using prior probabilities and Mahalanobis distances to recommend the most suitable trained machine-learning model for predicting lithology, thereby improving prediction accuracy and efficiency by comparing class-wise input probability distributions rather than global distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If global probability distribution comparison is used to select machine-learning models, then model selection can be performed, but memory efficiency is reduced and computational complexity increases
Solution Approach 1:
The patent segments the global probability distribution comparison into class-wise input probability distribution comparisons. Instead of comparing entire global distributions, the system divides the comparison into separate class-specific distributions, reducing the dimensionality and complexity of the computation while maintaining selection accuracy.
Solution Approach 2:
The patent extracts only the necessary class-wise input probability distributions from the full data sets, rather than processing complete global distributions. This extraction approach reduces memory requirements by storing only the essential distribution parameters needed for model selection.
2Reliability
If global probability distribution comparison is used to select machine-learning models, then model selection can be performed, but memory efficiency is reduced
Solution Approach 1:
The patent extracts only the necessary class-wise input probability distributions from the full data sets, rather than processing complete global distributions. This extraction approach reduces memory requirements by storing only the essential distribution parameters needed for model selection.
Solution Approach 2:
The patent segments the global probability distribution comparison into class-wise input probability distribution comparisons. Instead of comparing entire global distributions, the system divides the comparison into separate class-specific distributions, reducing the dimensionality and complexity of the computation while maintaining selection accuracy.
3Productivity
If machine-learning models are selected without considering probability distribution similarity, then classification can be performed quickly, but classification accuracy deteriorates when training and test data distributions differ
Solution Approach 1:
The patent performs preliminary action by calculating and storing class-wise input probability distributions from training data sets before model selection. This pre-computation allows for rapid comparison during model selection without sacrificing classification accuracy, as the distribution characteristics are already prepared.
Solution Approach 2:
The patent uses lightweight probability distribution representations (such as mean and covariance matrices) instead of storing complete data sets. These simplified distribution models require minimal memory and computation while effectively capturing the essential characteristics needed for accurate model selection.
Data Source
AI summary
According to some aspects, machine-learning models can be executed to classify a subsurface rock. Examples include training numerous machine-learning models using training data sets with different probability distributions, and then selecting a model to execute on a test data set. The selection of the model may be based on the similarity of each data point of the test data set and the probability distribution of each training class. Examples include detecting and recommending a pre-trained model to generate outputs predicting a classification, such as a lithology, of a test data set. Recommending the trained model may be based on calculated prior probabilities that measure the similarity between the training and test data sets. The model with a training data set that is most similar to the test data set can be recommended for classifying a physical property of the subsurface rock for hydrocarbon formation.