Multi-Head CNN for Seismic Rock Property Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional seismic inversion methods require significant time and are prone to human error due to reliance on assumptions, and they struggle to accurately predict rock properties away from drilled wells.
Innovation Solution
A data-driven deep learning system utilizing a multi-head Convolutional Neural Network (CNN) model that captures spatial and temporal relationships in 3D seismic data at different resolutions to estimate rock properties, reducing the need for human assumptions and improving prediction accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional seismic inversion methods are used, then human assumptions can guide the interpretation, but the process requires significant time and is prone to human error
Solution Approach 1:
The patent replaces the manual, mechanical process of conventional seismic inversion with an automated deep learning system. The CNN model automatically processes seismic data to generate rock property predictions, eliminating the need for manual interpretation while significantly reducing processing time from months to hours.
Solution Approach 2:
The patent performs preliminary training of the deep learning model using synthetic seismic data and well log data before actual deployment. This preliminary action creates a pre-trained system that can quickly make accurate predictions without requiring time-consuming manual adjustment during actual use.
2Reliability
If conventional seismic inversion methods are used, then the process can be completed with basic computational tools, but it struggles to accurately predict rock properties away from drilled wells
Solution Approach 1:
The patent transitions from traditional 1D seismic trace analysis to processing full 3D seismic volumes. By incorporating spatial relationships across multiple dimensions and using multi-head attention mechanisms, the model captures contextual information from surrounding areas, enabling accurate predictions away from well locations.
Solution Approach 2:
The patent implements a hierarchical deep learning architecture where convolutional layers extract local features that are then processed by higher-level layers to capture global patterns. This nested structure allows the model to build complex predictions from simpler intermediate representations, handling the complexity of predicting properties away from wells.
3Productivity
If deep learning models are used to reduce processing time, then prediction speed increases, but the model requires extensive training data and computational resources
Solution Approach 1:
The patent generates synthetic seismic data by copying and transforming existing well log data into synthetic seismic traces. This creates additional training data without requiring field surveys, allowing the model to be trained extensively on realistic data while avoiding the need for massive amounts of real seismic data.
Solution Approach 2:
The patent applies data augmentation techniques that transform training data by changing parameters such as adding noise, adjusting amplitude, and applying various filters. This increases the diversity and quantity of effective training data without requiring additional field measurements, enabling faster model training and deployment.
Data Source
AI summary
A system for estimating a rock property away from a well may include one or more hardware processors configured to access acquired three-dimensional (3D) seismic data that includes seismic traces from a 3D seismic survey of an area of interest. The system may also include a multi-head Convolutional Neural Network (CNN) model. The multi-head CNN model may include a plurality of kernels of various sizes for determining spatial and temporal relationships of the captured 3D seismic data at different resolutions. The multi-head CNN model may be trained to generate an estimated rock property value of a formation zone included in the area of interest, away from the well. The one or more hardware processors are further configured to update a drilling program for a production system based on the estimated rock property value. The drilling program may be executed on a computing device of the production system.