Neural Network Calibration for Prestack Seismic Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Prestack seismic inversion faces challenges in complex environments due to noise in large offset seismic data, loss of high-frequency details, and inherent uncertainties in seismic data and wavelet, which affect the accuracy and resolution of reservoir characterization.
Innovation Solution
A method using fully connected neural networks to calibrate prestack seismic inversion results, integrating seismic data, well log data, and geological attributes to produce more accurate and higher-resolution predictions of reservoir properties across the entire seismic survey region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If prestack seismic inversion is applied to complex environments, then comprehensive geological information can be obtained, but accuracy and vertical resolution deteriorate due to noise in large offset data and inherent uncertainties
Solution Approach 1:
The patent combines multiple data sources (seismic data, well log data) and multiple inversion results (P-wave velocity, S-wave velocity, density) into a unified neural network model. This integration allows the system to leverage the complementary strengths of different data types while compensating for their individual weaknesses, thereby maintaining comprehensive geological information while improving accuracy and resolution.
Solution Approach 2:
The neural network acts as an intermediary between the raw seismic inversion results and the final reservoir characterization. It processes and calibrates the inverted parameters by comparing them with well log data, effectively mediating the transition from uncertain inversion results to accurate reservoir properties. This intermediary processing step resolves the contradiction by filtering out noise and uncertainties while preserving comprehensive geological information.
2Loss of information
If large offset seismic data is used to obtain density properties, then more comprehensive reservoir information is obtained, but reliability deteriorates due to noise and lower signal quality
Solution Approach 1:
The patent implements a feedback mechanism where the neural network continuously compares seismic inversion results with well log data and adjusts the calibration accordingly. The well log data serves as ground truth feedback that validates and corrects the seismic-derived density properties. This feedback loop ensures that density information from noisy large offset data is reliably calibrated against high-quality well measurements, resolving the contradiction between comprehensive information and signal reliability.
Solution Approach 2:
The system transforms the unreliable density parameter derived from noisy seismic data into a calibrated parameter by applying neural network-based corrections. The calibration process changes the parameter's reliability by adjusting it based on the relationship between seismic attributes and well log measurements, thereby maintaining access to comprehensive density information while overcoming the noise and low signal quality issues.
3Area of stationary object
If seismic data and wavelet are used for inversion, then wide spatial coverage is achieved, but vertical resolution deteriorates due to band limiting and loss of high-frequency details
Solution Approach 1:
The patent segments the inversion process into multiple independent parameter estimations (P-wave velocity, S-wave velocity, density) rather than attempting to recover all frequency information simultaneously. The neural network processes each parameter separately and calibrates them individually against well log data. This segmentation allows the system to maintain wide spatial coverage from seismic data while achieving high vertical resolution through targeted calibration of each parameter, effectively bypassing the band-limiting constraint.
Solution Approach 2:
The patent transitions from a single-dimension frequency recovery problem to a multi-dimensional calibration problem by introducing well log data as an additional dimension. Instead of trying to recover lost high-frequency information directly in the frequency domain, the system uses the depth-domain well log measurements to calibrate the seismic inversion results. This dimensional transformation allows the system to achieve high vertical resolution without compromising the wide spatial coverage provided by the seismic data.
4Loss of information
If multiple parameters (P-wave velocity, S-wave velocity, density) are combined to derive fluid indicators, then more comprehensive reservoir characterization is achieved, but uncertainties compound and overall accuracy deteriorates
Solution Approach 1:
The patent applies preliminary calibration action to each individual parameter (P-wave velocity, S-wave velocity, density) before they are combined to derive fluid indicators. The neural network calibrates each parameter against well log data in advance, ensuring that the input parameters are already optimized and uncertainties are minimized. This preliminary calibration prevents the compounding of uncertainties when parameters are combined, allowing the system to achieve comprehensive fluid characterization while maintaining high overall accuracy.
Solution Approach 2:
The system transforms the individual seismic parameters into calibrated parameters through neural network processing, changing their uncertainty characteristics. By applying parameter-specific calibration transformations before combining them for fluid indicator calculation, the system ensures that uncertainties do not compound. The calibrated parameters have reduced and more uniform uncertainty distributions, enabling accurate fluid discrimination while maintaining comprehensive reservoir characterization.
Data Source
AI summary
An approach for calibrating prestack seismic inversion is provided. This method includes selecting various features from inverted elastic properties to generate reservoir properties; fully connected neural network models are used to learn the mapping between the features and ground truth data at well locations; and the prediction is applied to generate one or more final models for the reservoir characterization of the whole survey region.