Generative Triplet Model for Well Log Similarity

This model consists of a self-supervised triplet approach that levers the generative capability of GANs to evaluate well log similarity.

Execution

To execute the WellGT training software, follow this bash command:

python3 train.py --model <model> --dataset <dataset> --run <x>

Replace <model> and <dataset> with the desired model and dataset options respectively. Additionally, <x> indicates the number of the execution.

Available Models:

• WellGT (wellgt)(Ours)

Our model is showed below:

Baselines:

• Bootstrap Your Own Latent (byol)
• Dynamic Time Warp (DTW)(No training)
• Triplet Romanenkova (romanenkova)
• Variational AutoEncoder (vae)

Available Datasets:

• Force (force)
• Taranaki (taranaki)
• Force and Taranaki (public)

All execution parameters:

Command	Description	Default Value	Possible Values
--model	Name of the desired model.	'wellgt'	Any implemented model name
--dataset	Desired dataset for training and testing model.	'public'	Any implemented dataset name
--seq_size	Input sequence size (in data points, not meters/feet).	256	Positive integer
--interval_size	Minimum interval between the first value of two selected sequences.	128	Positive integer
--lr	Learning rate used to train models	0.0001	Positive float
--weight_decay	Weight Decay used by the optimizer	0.00001	Positive float
--batch_size	Batch size used to train the model	32	Positive integer
--epochs	Number of epochs used to train the model	500	Positive integer
--feature_size	Size of the embeddings created by the models	256	Positive integer
--z_size	Size of the latent vector z used by the generator in WellGT	100	Positive integer
--split_method	Method used for splitting data.	'cv'	Any implemented split method
--initial_margin	Initial margin value for the triplet loss margin parameter.	0.1	Positive float
--final_margin	Final margin value for the triplet loss margin parameter in the last epoch.	0.1	Positive float
--half_life	Number of epochs before the margin is half its initial value.	100	Positive integer
--swap	Swap hyperparameter value for the triplet loss.	True	True, False
--half_life	Number of epochs before the margin is half its initial value.	100	Positive integer
--save_model	Wether to save or not the trained model after training.	True	True, False
--save_dir	Directory used for saving the model.	model-checkpoints	Any existing path
--output_dir	Directory used for saving the model evaluation results.	results	Any existing path
--config_dir	Directory used for loading the config yml files.	configs	Any existing path
--verbose	Print the evolution of steps during execution	False	True, False
--run	Number of the code execution (useful for multiple runs using the same configs).	1	Positive integer

For example, to run the training code with the Variational Autoencoder on the Force dataset, the command would be:

python3 benchmark.py --model vae --dataset force

Configs

The 'configs' directory has '.yml' files that hold the necessary parameters to run the code properly.

data.yml:

This file contains necessary information to perform the methods.

Field	Description	Default Value	Possible Values
scaling_method	Method for scaling the data.	'standard'	'standard', 'minmax'
columns_used	Logs used to train models.	'GR', 'RHOB', 'NPHI', 'DTC'	Any list of existing logs
num_channels	Number of channels (logs). Length of columns_used.	4	Positive integer

<dataset>.yml files:

There are three YAML configuration files named <dataset>.yml, where <dataset> can be replaced with 'force', 'taranaki'. These files require modification in the data_dir parameter to specify the directory where the dataset is stored.

Example:

• force.yml: Configuration file for the 'force' dataset.
• taranaki.yml: Configuration file for the 'taranaki' dataset.

In each of these YAML files, locate the data_dir parameter and set its value to the directory path where the respective dataset is stored.

Methodology:

In our paper and technical report, we have justified our decision to use four key well logs in our approach: Gamma Ray, Density, Neutron, and Compressional Wave. We selected these logs because they are properties that are highly influenced by rock types and exhibit a degree of completeness, meaning they are less susceptible to missing data issues.

For evaluation, we use three experiments and each have their own metrics:

Sinusoidal Ranking

A sinusoidal curve dataset composed of sequences is generated through the formula:

The number of sequences is a hyperparameter selected when generating the dataset. The goal of similarity in this experiment: sinusoidal sequences with close frequencies, regardless of the phase, are similar, while sequences with different frequencies are dissimilar.

To evaluate the performance in this experiment, the following metrics are used:

• MAP@k (Mean Average Precision at k): Check the presence of top k relevant elements in top k predicted elements.
• MRR@k (Mean Reciprocal Rank at k): Give scores based on the position of the predicted sequences in the ground truth ranking.

Augmentation Ranking

As augmentation is our goal related to similarity, that is, we want the anchor and its augmentation to be similar, another approach to evaluate the models is by ranking the entire dataset with one augmentation of the query image in it. The goal is to establish where the models would rank the augmentation considering they are what is expected from well-log similarity. According to the augmentation's position, metrics are established to evaluate the performance of all the models.

To measure the performance, we use the following metrics:

• Accuracy@k: Checks if the augmentation is within the top k predicted elements.
• MRR (Mean Reciprocal Rank): Assigns scores based on the position of the augmentation in the predicted ranking.