Probabilistic Classifier for Petrophysical Classification with Augmented Well Log Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for petrophysical rock property calculation in reservoir characterization are limited by data sparsity and inaccuracies in predicting rock and fluid properties from seismic data, particularly in areas with few well data penetrations, leading to difficulties in accurately classifying lithofluid classes and estimating uncertainties.
Innovation Solution
The method involves generating augmented well log data by calibrating rock physics models using initial well log data, expanding porosity, saturation, fluid type, and mineralogy ranges to create a training dataset, and employing a probabilistic classifier, such as a Bayesian classifier, to calculate probability volumes and posterior probabilities for lithofluid classes, thereby improving the accuracy of reservoir characterization and uncertainty estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional well log data methods are used for petrophysical classification, then the process is simple, but data sparsity leads to inaccurate predictions of rock and fluid properties
Solution Approach 1:
The method performs preliminary actions by generating augmented well log data through rock physics model calibration before the actual classification task. This preprocessing step creates synthetic training data that expands the limited well log information, thereby improving prediction accuracy without requiring additional physical measurements.
Solution Approach 2:
The invention creates copies of well log data by generating synthetic well log realizations through probabilistic classification and rock physics modeling. These synthetic copies augment the sparse original data, providing sufficient training material for accurate petrophysical property prediction while maintaining the statistical characteristics of the original data.
2Reliability
If more well data penetrations are obtained to improve classification accuracy, then prediction reliability improves, but the cost and complexity of acquiring additional data increases
Solution Approach 1:
The method introduces rock physics models and probabilistic classifiers as intermediary components that bridge the gap between sparse well log data and reliable petrophysical predictions. These intermediaries process and transform the limited input data into robust classification results without requiring additional physical data acquisition infrastructure.
Solution Approach 2:
The invention changes parameters by transforming deterministic well log measurements into probabilistic distributions through Bayesian classification. This parameter transformation allows the system to quantify uncertainty and improve reliability by expressing predictions as probability volumes rather than single deterministic values.
3Loss of information
If deterministic classification methods are used, then the process is computationally efficient, but uncertainty estimation cannot be provided
Solution Approach 1:
The method transitions from static deterministic classification to dynamic probabilistic classification that adapts to data variability. By using Bayesian classifiers that compute probability distributions rather than fixed class assignments, the system dynamically captures uncertainty information while maintaining computational tractability through efficient sampling and aggregation methods.
4Adaptability or versatility
If rock physics models are calibrated to expand data ranges, then training dataset robustness improves, but the complexity of model calibration increases
Solution Approach 1:
The invention makes the rock physics models multi-functional by using them for both physical property prediction and synthetic data generation. This universal application allows the same calibrated models to serve dual purposes: predicting petrophysical properties and generating augmented training data, thereby expanding data range coverage without proportionally increasing calibration complexity.
Data Source
AI summary
Techniques and systems to provide increases in accuracy of property determination of a formation. The techniques include receiving initial well log data, generating augmented well log data including the initial well log data and modeled well log data based on the initial well log data, modifying the augmented well log data to generate a training dataset, training a probabilistic classifier utilizing the training dataset, calculating a probability volume for each lithofluid class of a set of predetermined lithofluid classes utilizing the probabilistic classifier, outputting the probability volume for each lithofluid class of the set of predetermined lithofluid classes as a respective probability of an occurrence of a type of lithofluid class in a reservoir, calculating a posterior probability based on the probability volume for a first lithofluid class of the set of predetermined lithofluid classes, and outputting the posterior probability as a probability of a property of the reservoir.