Physics-driven Deep Learning Pixel-Based Inversion of Logging-While-Drilling in Anisotropic Formation

Citation Author(s):
Ning
Li
Henan Polytechnic University
Submitted by:
Ning Li
Last updated:
Sun, 12/03/2023 - 08:35
DOI:
10.21227/cwgf-g832
License:
0
0 ratings - Please login to submit your rating.

Abstract 

We propose a coupled physics-driven and data-driven algorithm to improve standard deep learning workflow. In order to evaluate the proposed method, a 2.5D geological model including dip, fault and anisotropic formation is considered.  Comparing the inversion imaging performance of the proposed physics-driven method with the traditional classical residual network (Resnet), it shows a significant improvement in resistivity accuracy.

The folder of Dataset File contains six sub-folders. The Fault classification folder is the data File of fault identification workflow, which contains the amplitude and phase data sets and labels of fault identification. Fault identification workflow training loss function source file.

Loss_function folder is the source file of loss function of inversion workflow, in which cnn_all_train_loss.txt is the source file of loss function of training set without fault based on Resnet network. cnn_all_val_loss.txt is the source file of fault free test set loss function based on Resnet network, and cnn_fault_train_loss.txt is the source file of fault training set loss function based on Resnet network. cnn_fault_val_loss.txt is the source file of fault test set loss function based on Resnet network. The last four are physics-driven loss function source files that are represented the same as the Resnet network.

Noise_result folder adds noise inversion results, including noiseless, strong and weak noise inversion result source files of physics-driven continuous models for noise analysis as well as real model source files.

The Test_result folder is the source file for test results based on the physics-driven and traditional Resnet network graben segmentation model.

The Training_sets folder is the input data set file for workflow training inversion, where layer_traject.txt is the trajectory coordinate file of LWD. train_layer_amplitude.txt is the forward amplitude response file of the model without fault, train_layer_phase.txt is the forward phase response file of the model without fault, train_fault_amplitude.txt is the forward amplitude response file of the model with fault. train_fault_phase.txt is a forward phase response file with fault model. Qzz_layer_amplitude.txt is the amplitude ratio response file after forward modeling without fault, and qzz_layer_phase. txt is the phase difference response file after forward modeling without fault. Qzz_fault_amplitude.txt is the amplitude ratio response file after the forward modeling of fault model, and Qzz_fault_phase.txt is the phase difference response file after the forward modeling of fault model. train_layer_resistivity_h.txt is the horizontal resistivity file of the non-fault model, and train_layer_resistivity_v.txt is the vertical resistivity file of the non-fault model. train_fault_resistivity_h.txt indicates the horizontal resistivity file with the fault model, and train_fault_resistivity_v.txt indicates the vertical resistivity file with the fault model.

The True_model folder is the actual file of the graben split model, containing the resistivity file, LWD trace file, and forward response (amplitude and phase) file for each model.

The Script folder contains 12 Python file scripts, classified.py is the network training script of fault identification workflow; classification_loss_plot.py drawing script for network training loss function image of fault identification workflow; classification_plot.py probabilistic bar plot script containing faults to identify model for fault identification workflow; classification_test.py identifies workflow test script for fault; train_physics.py is a physics-driven inversion workflow training script. train_tradition.py is an inversion workflow training script based on Resnet network. mt1dfwd.py is a script for calculating the forward response of geological models.py and cluster_tradition.py are R2 evaluation image drawing scripts based on physics-driven and Resnet networks, respectively. Cluster_plot_ms.py is the script for drawing the RMS bar chart. plot_result.py script for drawing inversion result images from inversion result files; test.py is a script to test the network.

Instructions: 

We propose a coupled physics-driven and data-driven algorithm to improve standard deep learning workflow. We open source this workflow data and code in IEEE DataPort. In order to evaluate the proposed method, a 2.5D geological model including dip, fault and anisotropic formation is considered. Comparing the inversion imaging performance of the proposed physics-driven method with the traditional classical residual network (Resnet), model tests show a significant improvement in resistivity accuracy.