Synthetic Well Data Generation for Seismic Inversion Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional seismic inversion methods require extensive and costly physical exploration to gather sufficient high-quality training data, making them inefficient and expensive, especially in frontier exploration areas with limited drilled wells.
Innovation Solution
The method involves dividing well data into segments, generating synthetic well data, and using random sampling with replacement to create improved training data for seismic inversion models, which includes pairings of seismic attributes and subsurface characteristics, enabling the training of effective seismic inversion models without extensive physical exploration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physical exploration (drilling wells) is performed to gather training data, then data quality and quantity improve, but cost increases significantly
Solution Approach 1:
The patent creates synthetic well data that copies the essential characteristics and statistical properties of real well data without requiring actual physical drilling. By generating artificial training datasets that mimic real subsurface configurations, the system achieves sufficient training data quality at minimal cost, directly resolving the contradiction between data reliability and exploration cost
Solution Approach 2:
The patent systematically varies key subsurface parameters (porosity, permeability, lithology, fluid saturation) in synthetic well data generation to create diverse training scenarios. This parameter variation approach ensures comprehensive coverage of possible subsurface conditions while avoiding the high costs of multiple physical drilling operations, thus improving training data quality without proportional cost increase
2Measurement precision
If more wells are drilled to gather sufficient training data, then machine learning model accuracy improves, but time and resources increase
Solution Approach 1:
The patent performs preliminary synthetic well data generation and model training in advance, creating a library of training datasets and pre-trained models before actual seismic inversion is needed. This preliminary action eliminates the need for time-consuming physical drilling and data collection during the actual inversion process, thereby improving inversion accuracy without proportional time loss
Solution Approach 2:
By copying real well data patterns and statistical properties into synthetic datasets, the system achieves sufficient training quality without requiring actual drilling operations. This copying approach provides all necessary training data instantaneously, improving model accuracy while eliminating the time-consuming nature of physical exploration
3Loss of information
If extensive physical exploration is conducted to obtain training data, then subsurface understanding improves, but cost and environmental impact worsen
Solution Approach 1:
The patent creates virtual copies of subsurface environments through synthetic well data, capturing all necessary geological information without physical intrusion. This copying methodology achieves complete subsurface understanding while eliminating environmental disturbance associated with drilling operations, directly resolving the contradiction between information gain and environmental harm
Solution Approach 2:
The patent introduces synthetic data generation algorithms as an intermediary between the need for subsurface knowledge and the avoidance of physical drilling. This intermediary process translates real geological patterns into virtual training data, providing comprehensive subsurface understanding without direct environmental impact from drilling activities
Data Source
AI summary
Well data (e.g., well log) may be divided into multiple segments, and different samplings of data in the individual segments may be performed to increase the amount of data that is used to train a seismic inversion model. Synthetic well data may be generated from real well data to increase the amount of well data from which sampling is performed.