Physics-Constrained Machine Learning for Sonic Log Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Well logs often contain missing or poor-quality data due to malfunctioning logging tools or incorrect recording, which can lead to uncertainties affecting subsequent drilling operations, limiting the accuracy of mechanical property calculations essential for well planning.
Innovation Solution
A physics-constrained machine learning (PCML) model is trained using well logs data to predict and correct sonic logs and mechanical properties, incorporating a breakout model to validate and update the predictions based on field data, ensuring accurate mechanical earth modeling for optimized drilling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If traditional machine learning models are used to predict sonic logs, then the model can learn from data patterns, but the predictions may produce non-physical values that violate rock mechanics principles
Solution Approach 1:
The patent transforms the machine learning output from raw predictions to physically constrained predictions by changing the parameter space. The loss function incorporates physical constraints (Poisson's ratio between 0.04-0.3, Young's modulus positivity) to ensure predictions remain within physically valid ranges, thus resolving the contradiction between automation and reliability.
Solution Approach 2:
The patent introduces physics-based rock mechanics equations as an intermediary layer between the machine learning model and the final predictions. The mechanical properties (Poisson's ratio, Young's modulus) serve as mediators that constrain the sonic log predictions to satisfy physical laws, ensuring reliability while maintaining predictive capability.
2Loss of information
If well logs are collected using physical logging tools, then detailed sequential data can be obtained, but the data may be missing or poor quality due to tool failures or malfunction
Solution Approach 1:
The patent uses a feedback mechanism where the machine learning model predicts missing sonic log values based on other well log measurements (gamma ray, density, neutron porosity). The model learns from available data patterns and continuously refines predictions to compensate for missing or poor quality data, thus recovering information loss while maintaining reliability.
Solution Approach 2:
The patent makes the machine learning model multi-functional by enabling it to predict multiple sonic log parameters (compressional and shear wave velocities) simultaneously from various input well logs. This universal approach allows the system to handle different types of data loss scenarios and recover comprehensive sonic information even when specific logging tools fail.
3Reliability
If physics constraints are incorporated into the machine learning model, then physically valid outputs are ensured, but the model complexity increases
Solution Approach 1:
The patent manages model complexity by parameterizing the physical constraints through a loss function rather than building complex physical simulation models. By defining simple boundary conditions (Poisson's ratio range, Young's modulus positivity) in the optimization objective, the patent achieves physical validity without substantially increasing model structural complexity.
Solution Approach 2:
The patent replaces complex mechanical rock physics simulations with a data-driven machine learning model constrained by simplified physical principles. Instead of using computationally intensive mechanical models, the patent uses a neural network with physics-informed loss functions, substituting complex mechanical systems with a more efficient physics-constrained statistical approach.
Data Source
AI summary
A computer-implemented method may include obtaining well logs data pertaining to a well of interest. The method may further include training a physics-constrained machine learning (PCML) model using the obtained well logs data as inputs. The method may further include outputting one or more sonic logs and mechanical properties of interest determined by using the trained PCML model and the obtained well logs data for the well of interest. The method may further include updating the determined sonic logs and mechanical properties of interest based on a breakout model and field breakout data for the well of interest. The method may further include outputting the final sonic logs for the well of interest. The method may further include determining one or more mechanical properties for well planning based on the final sonic logs for the well of interest.