Bin-Dependent ML Prediction of Seismic First Arrivals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing seismic data processing methods face challenges in accurately determining first arrivals due to the presence of outliers, which can introduce errors and reduce the precision of seismic data processing operations, particularly in the context of hydrocarbon reservoir exploration.
Innovation Solution
A method and system utilizing a trainable machine-learning (ML) network to process seismic datasets, where the network is trained with initial first arrivals determined from a representative portion of the seismic dataset, allowing it to predict first arrivals with reduced outliers and improve the quality of seismic data processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional seismic data processing methods are used to determine first arrivals, then the processing can be performed with simple algorithms, but the presence of outliers introduces errors and reduces measurement precision
Solution Approach 1:
The patent replaces traditional mechanical/algorithmic signal processing with a machine learning-based system. The ML network is trained on seismic data to predict first arrivals, substituting conventional pickup algorithms with a data-driven model that can handle complex patterns and outliers more effectively, thereby improving measurement precision without requiring overly complex processing infrastructure
Solution Approach 2:
The patent creates a digital model (copy) of seismic wave propagation behavior through machine learning training. By training the network on representative seismic data, the system creates a virtual replica that can predict first arrivals without physically analyzing every complex interaction, allowing for high-precision measurements while keeping the actual processing system relatively simple
2Reliability
If machine learning networks are trained on representative portions of seismic data, then the network can predict first arrivals with reduced outliers, but the training process requires partitioning data into bins and subsets
Solution Approach 1:
The patent divides the seismic data into discrete bins based on spatial or temporal characteristics, and further partitions these bins into training subsets. This segmentation allows the complex training process to manage data in manageable chunks, improving reliability by ensuring the network learns from representative samples while keeping the training process organizationally simple through systematic binning
Solution Approach 2:
The patent applies different training strategies to different bins and subsets based on their local characteristics. By identifying which portions of the seismic data are most representative or problematic, the system can focus training effort where it matters most, improving overall reliability without requiring uniform complex processing of all data
Data Source
AI summary
Examples of methods and systems are disclosed. The methods may include receiving a seismic dataset, wherein the seismic dataset comprises a plurality of time-space waveforms, and forming a training waveform set from a subset of the plurality of time-space waveforms. The methods may also include generating a plurality of training subsets from the training waveform set and determining a plurality of initial first arrivals based on the plurality of training subsets. The methods may further include forming a training dataset, wherein the training dataset comprises an input training dataset and an output training dataset, wherein the input training dataset is based on the plurality of training subsets and the output training dataset is based on the plurality of initial first arrivals, and training, using the training dataset, a machine-learning (ML) network to predict the output training dataset, at least in part, from the input training dataset.