Prediction of missing well logs using ML

Well Logs NN Predictor

LogNNPredictor: Well log prediction using machine learning

LogNNPredictor is a two-part tool designed to help geroscientists predict well logs using machine learning. In Part 1, users upload LAS files and train a neural network model by selecting input and target logs, with customizable training parameters. In Part 2, the trained model is applied to new LAS files to predict well logs, generating new files with the predicted values and visual comparisons of actual vs predicted logs. This tool streamlines the process of building and applying predictive models for well log data.

slipExplore

What these apps do

LogNNPredictor is a two-part tool designed to help geroscientists predict well logs using machine learning. The process is divided into two steps:

Part 1 - Training: In the first part, you upload LAS files containing well log data from your local drive. You select the target log you want to predict and the input logs to train the model. You can customize parameters such as the number of epochs, neurons, and learning rate. The tool then trains a neural network model to predict the selected target log based on the input logs. Once training is complete, the model is saved for future use.

Part 2 - Modeling: In the second part, you apply the trained model from Part 1 to new LAS files. The tool predicts the target log values based on the input logs in the new data. It also generates new LAS files with the predicted logs and provides visual comparisons between the actual and predicted values, if available.

Together, these two steps allow you to train a model on existing well log data and use it to make predictions on new datasets, helping you fill gaps or forecast future log behavior.

Click here to download a test dataset and detailed instructions to help you get started before using your own data.

You must have a Google account to run these apps. The apps neither access any data in your local drive other than the one you indicate, nor we store or share your data in any form. For more information, read our Privacy Policy and Terms of Service. For additional system & browser requirements, see below.

To run the apps, click the buttons below. For the latest version, make sure to run them from these buttons. The apps are not designed to work on mobile devices.

Run Part 1 -Training

Run Part 2 - Modeling

Training

LogNNPredictor: Part 1 - Training

The first step of the log prediction process if the generation of the training model used to predict logs in Part 2. Below is a description of the input data, parameters, and outputs.

Input data (LAS files)

LAS files( in local drive) containing well log data used for training the prediction model. All LAS files must contain the same curves, in the same order.
Multiple LAS files can be uploaded.
A target log (the log you want to predict) and input logs (logs used to predict the target) are selected from the uploaded data.

Input parameters

We've selected default parameters that work well for many situations but you may need to experiment with them to enhance the results in your particular case.

Null Value: Value used to indicate missing data in any raw of data in the LAS files (e.g., -999.25). Rows with a null value in any logs are not used for training. Read below on how to run the app to predict missing sections of well logs (due to washouts, for instance).
Target Log: The specific log to be predicted (e.g., gamma ray log).
Input Logs: Logs used as inputs to the model for training (e.g., density, sonic, or neutron logs). The target log will be a combination of the input logs.
Number of Epochs: Defines how many times the model will go through the training data. Higher epochs may increase accuracy.
Number of Neurons: Specifies how complex the model will be. More neurons may help capture complex relationships but could risk overfitting.
Batch Size: Controls how many data points are used at a time during training. Larger batches may speed up training but require more memory.
Learning Rate: Adjusts how quickly the model adapts during training. A lower rate might lead to more precise training but takes longer.
Optimizer: Method used to adjust the learning process. Options include Adam (adaptive method) and SGD (Stochastic Gradient Descent).
Normalization Method: Adjusts the scale of the input data to improve model performance. Two options: 1) Standard: Removes the mean and scales data to unit variance. 2) Min-Max: Scales values between a minimum and maximum value.
Validation Split: Defines how much of the data will be used for validating the model, typically a percentage (e.g., 20%).

Output files

In Part 1 of LogNNPredictor, three output files are generated:

Model Configuration (JSON) stores the model’s structure, including layers and neurons.
Model Weights (BIN) contains the trained parameters (weights) optimized during training.
Training Data (JSON) includes metadata such as input logs, target log, and normalization values, as well as training parameters like epochs and batch size.

Click on "Save Model Files (3)" to download the files to your local drive. These files are crucial for replicating the trained model and ensuring consistent predictions in Part 2.

LogNNPredictor: Part 2 - Modeling

This app is the second step in the well log prediction process. After training a model with the LogNNPredictor: Part 1 - Training, you can now apply the trained model to new LAS files to predict well logs. Below is a description of how the app works and what it outputs.

Input data

Training Data: The JSON file generated in Part 1 contains important parameters such as input logs, target log, and normalization values used during the training process.
Model Configuration: The JSON file containing the architecture and settings of the trained model from Part 1.
Model Weights: A .bin file storing the trained weights from Part 1, allowing the model to make accurate predictions.
LAS Files: LAS files (in local drive) containing input well logs where the trained model will be applied to predict target logs. All files must contain the same curves used in Part 1, in the same orders, except for some files that may contain the target curve too.

Output parameters and files

.Output LAS Files Suffix: A custom suffix that will be appended to the name of the newly generated LAS files containing predicted values.

Click on "Apply model" to make the predictions. If you like to save the the predicted log, click on "Save LAS Files" to download the LAS file(s) with the predicted curve. Depending on the "Download" settings of your browser, they may go directly to your local "Downloads" folder or you can change these settings to better suit yours needs.

Other uses of the LogNNPredictor app:

Prediction of missing sections of logs.

This free web app is designed to predict fully missing logs in wells where other logs are available. However, it does not directly handle cases where the log is only partially missing in sections along the well (such as due to washouts). Here’s a workaround for such situations:

If you want to fill gaps in a density log but plan to use the same well for training, you should generate two LAS files. The first file is used in Part 1 (Training) and should contain both the input logs (e.g., NPHI, GR, and Resistivity) and the target log (Density). The second file is for Part 2 (Modeling) and contains only the input logs (NPHI, GR, and Resistivity). The output of Part 2 will provide a density estimate for the entire well. You can then use this estimate to fill in the missing sections of the density log where needed.

Facies estimation.

If the target log represents discrete facies (such as 0 and 1), the app can still be used for facies prediction. In this case, the predicted log will output a continuous probability distribution of the facies of interest. You’ll need to convert this continuous probability back into discrete values by applying cutoffs. These cutoffs should be calibrated to match the original facies data as closely as possible, ensuring accurate classification. Then, you can apply the same rules in wells where no facies logs are available.

System & browser requirements for using the LogNNPredictor app:

Google account required:

Users must be logged into a Google account to use the app, as it interacts with Google Services.

JavaScript must be enabled:

Ensure that JavaScript is enabled in the browser to allow the app’s scripts and interactive elements to function properly (e.g., TensorFlow.js and Plotly.js for model training and graph visualization).

Supported browsers:

The app works best with the latest versions of modern browsers like Google Chrome, Mozilla Firefox, Microsoft Edge, Safari (on Mac).

Clear cache / Restart browser:

If you are running the app multiple times with different wells, you may experience performance issues. It's recommended to:

Clear your browser cache regularly.

Restart the browser after a few runs, as memory usage might build up, slowing down performance over time.

Recommended memory:

While there are no strict memory requirements, having at least 8 GB of RAM is recommended for smooth operation, especially when working with large datasets or running the app multiple times.

Internet connection:

A stable internet connection is necessary to load external JavaScript libraries (e.g., TensorFlow.js) and interact with Google services.

Browser extensions:

Disable or whitelist the app in any script-blocking extensions (such as NoScript, uBlock Origin, or similar) as these may interfere with the app's functionality.

Modeling