Subsurface Time-Lapse Property Estimation Using Synthetic Seismic Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating time-lapse property changes in subsurface volumes, such as cross-correlation and inversion, often produce anomalous time shifts and lose spatial resolution due to noise and require manual tuning, leading to false 4D seismic interpretation.
Innovation Solution
A computer-implemented method using a backpropagation-enabled model, trained with synthetic seismic data, directly maps baseline and monitor seismic traces to time shifts without the need for tuning window lengths, employing supervised learning with ground truth labels to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If trace-by-trace cross-correlation is used to compute time shifts, then the method is reliable and easy to implement, but it produces anomalous time shifts when seismic amplitudes change, leading to false 4D interpretation
Solution Approach 1:
The patent replaces the mechanical cross-correlation method with a machine learning model that has been trained on synthetic seismic data. The ML model learns the complex relationship between baseline and monitor traces and directly predicts time shifts, avoiding the mathematical limitations of cross-correlation when amplitudes change. This substitution transforms a deterministic mathematical operation into a data-driven predictive system.
Solution Approach 2:
The patent performs preliminary training of the machine learning model using synthetic baseline-monitor trace pairs with known time shifts before applying the model to real data. This pre-training phase allows the model to learn the underlying patterns and relationships in seismic data, enabling it to accurately predict time shifts in actual monitoring scenarios without being affected by amplitude changes.
2Measurement precision
If a narrow window is used in cross-correlation, then noise impact is reduced, but time shifts are greatly impacted by noise
Solution Approach 1:
The patent uses synthetic copies of seismic data with known ground truth time shifts to train the machine learning model. By learning from these synthetic examples that replicate various noise conditions and amplitude changes, the model develops robustness to noise while maintaining precision in time shift estimation, without needing to manually tune window parameters.
3Reliability
If a broad time window is used in cross-correlation, then noise impact is reduced, but time shifts tend to be spread out and lose spatial resolution
Solution Approach 1:
The patent employs a dynamic machine learning model that can adaptively process seismic traces of varying lengths and characteristics. Unlike the static cross-correlation window approach, the ML model dynamically learns the appropriate temporal relationships from training data, maintaining spatial resolution while being robust to noise through its learned representations rather than fixed window parameters.
4Reliability
If inversion method is used to estimate time shifts, then time shifts anomalies can be avoided, but the setup becomes complicated and is limited by local minimum issues
Solution Approach 1:
The patent replaces the complex iterative inversion process with a trained machine learning model that directly predicts time shifts from baseline and monitor traces. This substitution eliminates the need for complicated inversion setup, parameter tuning, and iterative optimization, while avoiding local minimum issues by using a pre-trained model that has already converged to optimal solutions during training on synthetic data.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A backpropagation enabled model is trained for estimating time-lapse property changes of a subsurface volume. Synthetic models of the subsurface volume are generated, with pre-determined property changes before and after a time lapse. These models are used to compute baseline-monitor pairs of synthetic seismic traces before and after the time lapse, wherein the baseline synthetic traces are computed from the synthetic model before the time lapse and the monitor synthetic traces are computed from the synthetic model after the time lapse. A ground truth 4D attribute characterizing the time-lapse property changes in the synthetic models is defined, and a backpropagation enabled model is trained by feeding the baseline-monitor pairs of synthetic seismic traces and the corresponding ground truth 4D attribute. The thus obtained trained backpropagation enabled model can be used to estimate time-lapse property changes of the actual subsurface Earth volume from actual baseline-monitor pairs of seismic traces.