Machine Learning Lithology Probability Estimation for AVA Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Determining lithology probabilities within a subsurface volume of interest is challenging due to limited subsurface data, making accurate amplitude versus angle of incidence (AVA) inversion difficult, especially with limited spatial sampling of well locations.
Innovation Solution
A system and method that uses machine learning techniques to estimate lithology probabilities as prior information in stochastic AVA inversion, combining seismic and well log data to improve interpolation and inversion accuracy by training a model with textural attributes from seismic and well data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional interpolation methods are used with limited well data, then the inversion process can proceed, but the accuracy of lithology probability estimates deteriorates
Solution Approach 1:
The patent introduces machine learning models as intermediary components that process seismic data and well log data to generate probabilistic lithology estimates. These models act as mediators between the limited well data and the broader subsurface volume, enabling accurate predictions without requiring extensive direct measurements across the entire volume.
Solution Approach 2:
The patent creates virtual copies of subsurface properties by using machine learning models to predict lithology probabilities at locations where no physical well data exists. The models learn patterns from well log data and replicate these patterns across the seismic volume, effectively copying known geological characteristics to unknown regions.
2Measurement precision
If more well locations are added to improve spatial sampling, then lithology estimation accuracy improves, but the cost and complexity of data acquisition increases
Solution Approach 1:
The patent makes the existing well log data and seismic data serve multiple functions: they are used both for direct lithology identification at well locations and as training data for machine learning models that predict lithology across the entire subsurface volume. This multi-functionality eliminates the need for additional well locations while achieving comprehensive coverage.
Solution Approach 2:
The patent transforms the approach from direct measurement-based estimation to model-based probabilistic estimation. By changing the parameter representation from deterministic well log values to probabilistic predictions from trained models, the system achieves better spatial coverage without additional physical measurements.
3Device complexity
If deterministic inversion is used, then the inversion process is simpler, but the ability to capture uncertainty and provide probability distributions deteriorates
Solution Approach 1:
The patent performs preliminary probabilistic estimation using machine learning models before conducting the inversion process. The models generate initial probability distributions for lithology types, which then serve as informed priors for the inversion. This preliminary action ensures that uncertainty information is captured and propagated through the entire workflow.
Solution Approach 2:
The patent implements a feedback loop where machine learning models continuously refine lithology probability estimates based on inversion results and seismic data. The models are trained on well log data, make predictions, receive feedback from inversion outcomes, and improve their predictions iteratively, maintaining both simplicity and uncertainty characterization.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods for training a model that uses probabilities of lithologies as prior information in an inversion are disclosed. Exemplary implementations may: obtain training data, the training data including (i) subsurface map data sets, and (ii) known lithologies; obtain an initial seismic mapping model; generate a conditioned seismic mapping model by training the initial seismic mapping model; store the conditioned seismic mapping model; obtain a target subsurface map data set; apply the conditioned seismic mapping model to generate a classified lithology map data set; apply an inversion to the classified lithology map data set to generate volumes of lithologies; generate an image that represents the volumes of lithologies; display the image.