Attentive Neural Property Modeling for Sparse Geological Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques like generative adversarial networks (GANs) struggle to provide a quantitative estimate of uncertainty in subsurface property modeling from sparse well measurements, limiting the reliability of geological models and increasing exploration risks.
Innovation Solution
The use of an attentive neural process (ANP) model that incorporates an attention layer to focus on relevant context locations, enabling accurate prediction of geological properties and uncertainty estimation by generating predicted mean and variance images from sparse context inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If GANs are used for property modeling, then realistic patterns are generated, but quantitative uncertainty estimation is not provided
Solution Approach 1:
The model segments the prediction task into two distinct components: mean prediction and uncertainty estimation. The neural process architecture separates these functions by having the network output both the mean values and variance values independently, allowing each to be optimized for its specific purpose without interfering with the other.
Solution Approach 2:
The patent introduces an intermediary mechanism - the neural process framework - that bridges GANs and Kriging. This framework adopts the GAN's ability to generate realistic patterns while incorporating Kriging's uncertainty estimation capability, creating a hybrid model that combines strengths of both approaches.
2Quantity of substance
If sparse well measurements are used for modeling, then data collection cost is reduced, but model accuracy deteriorates
Solution Approach 1:
The model creates virtual copies of the sparse measurements through the neural process framework. By learning the underlying distribution from limited well data, the model generates synthetic property values at unsampled locations that replicate the statistical characteristics and spatial patterns of the original data, effectively amplifying the information content of sparse measurements.
Solution Approach 2:
The patent transforms the problem from direct interpolation to distributional prediction. Instead of predicting single deterministic values, the model predicts the full probability distribution (mean and variance) of property values, allowing it to leverage sparse measurements more effectively by capturing uncertainty and variability patterns.
3Stability of the object's composition
If domain knowledge is incorporated into modeling, then geological consistency is improved, but model complexity increases
Solution Approach 1:
The model performs self-service by automatically learning geological patterns and constraints from the training data without requiring explicit domain knowledge encoding. The neural process framework learns the inherent structure and relationships in the well measurements, allowing the model to maintain geological consistency through data-driven pattern recognition rather than rule-based constraints.
Data Source
AI summary
A method implements property modeling using attentive neural processes. The method includes receiving sparse context input comprising a plurality of context locations corresponding to a plurality of geological property values for a geological property and selecting a plurality of target locations in a space of the plurality of context locations. The method further includes generating a predicted mean image for the geological property by an attentive neural process model using the plurality of target locations and the sparse context input and presenting the predicted mean image.