Hybrid Analytic and Machine Learning Petrophysical Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current seismic data processing methods, such as surface-wave inversion, are computationally intensive and sensitive to noise, making it difficult to accurately determine petrophysical property values within underground structures, especially when dealing with large-scale problems and noisy data.
Innovation Solution
A hybrid approach combining analytic inversion and machine learning using a deep neural network (DNN) is employed, where a sparse grid of 1D velocity models obtained through analytic solutions is used to train the DNN, allowing it to predict property values across the entire underground structure, thereby reducing computational effort and improving robustness to noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If analytic inversion methods are used to estimate property values from seismic data, then measurement precision is improved, but productivity deteriorates due to computationally intensive iterative inversions required at each spatial grid
Solution Approach 1:
The patent pre-computes a comprehensive lookup table containing property values for various seismic data configurations before actual processing. This preliminary action stores pre-calculated results that can be quickly retrieved during inversion, avoiding repeated computationally intensive calculations at each spatial grid point while maintaining estimation accuracy.
Solution Approach 2:
The patent creates a simplified copy of the complex inversion problem by pre-calculating and storing results in a lookup table. Instead of performing full iterative inversions during processing, the system copies pre-computed property value relationships and uses them for rapid estimation, significantly improving computational efficiency.
2Measurement precision
If analytic inversion is performed at each spatial grid independently, then measurement precision is improved through local optimization, but reliability deteriorates because the local sensitivity cannot see the noise distribution
Solution Approach 1:
The patent merges multiple local inversions by combining the pre-computed lookup tables from different spatial grids. This integration allows the system to leverage information from neighboring grids, enabling noise distribution analysis across the entire dataset while maintaining local optimization benefits, thus improving reliability and noise robustness.
Solution Approach 2:
The patent implements feedback mechanisms where inversion results from one grid point inform and constrain inversions at neighboring points. By using the pre-computed lookup tables collectively, the system feeds back noise characteristics and property value trends across the survey area, allowing each local inversion to benefit from global noise distribution information.
3Productivity
If a comprehensive lookup table is pre-computed for all possible seismic data configurations, then productivity is improved through faster property value estimation, but device complexity increases due to large memory requirements
Solution Approach 1:
The patent segments the comprehensive lookup table into smaller, manageable subsets organized by seismic data configuration types. Instead of storing one monolithic table, the system divides the pre-computed data into multiple smaller tables that can be loaded and processed in segments, reducing peak memory requirements while maintaining fast access speeds during inversion.
Solution Approach 2:
The patent applies local quality by loading only the relevant portion of the lookup table corresponding to the current seismic data being processed. Rather than having the entire comprehensive table in memory simultaneously, the system accesses only the local subset needed for the current grid point or data configuration, optimizing memory utilization.
Data Source
AI summary
Property values inside an explored underground subsurface are determined using hybrid analytic and machine learning. A training dataset representing survey data acquired over the explored underground structure is used to obtain labels via an analytic inversion. A deep neural network model generated using the training dataset and the labels is used to predict property values corresponding to the survey data using the DNN model.