ML for rock property prediction using seismic+wells

[Bandurao Naik]

Dear All,

How to parameterize the ML process and to chose the ML algorithm for better results. Are we going to try on trial and error basis?

Thanks...

SBR Naik

[Paul de Groot]

Dear Mr. Naik,

These are good questions with no unique answer. Let me share a few personal observations:

1. ChimneyCube: I prefer the original MLP network that is trained on the default attribute set over the new Seismic classification (image-to-point) workflow that uses dGB LeNet Classifier (a CNN model). The MLP result is much more fuzzy, which helps in the interpretation;
2. Seismic classification workflows (image-to-point) in general: easy to create training sets (pick or spray locations) but too slow for operation use in the application phase. If possible try instead to create a training set for a seismic segmentation (image-to-image) workflow;
3. Seismic Segmentation and Seismic Regression workflows: Use the dGB U-Net for 2D and 3D shapes. Typically, I try to create 10,000+ training examples. In most cases you can control the number of examples by increasing the overlap and/or by using data augmentation. In some cases training 15 epochs is sufficient, in other cases you may need 60 or more. Use the TensorBoard error graphs to decide when to stop or use Resume training for another number of epochs. I prefer PyTorch over TensorFlow because it uses less (GPU) memory, hence you can train larger shapes and larger batches. If your computer supports this, you can use mixed precision in training. This saves a lot of memory and speeds up training time. Shape sizes depend on your computer capacity. Typical sizes are 128x128, 256x256, 128x128x128;
4. Well Log Predictions: I prefer XGBoost on Random Forest as it gives the best predictions. The defaults are a good starting point.

I hope this helps.

Best regards,

Paul.

--

Paul de Groot

Special Adviser

-----------------------------------------------------------------------------------------------------------------------   dGB Earth Sciences
Phone: +31 53 4315155 E-mail: paul.d...@dgbes.com Internet: dgbes.com

----------------------------------------------------------------------------------------------------------------------- 

--
You received this message because you are subscribed to the Google Groups "OpendTect Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to users+un...@opendtect.org.
To view this discussion visit https://groups.google.com/a/opendtect.org/d/msgid/users/CABbBd9r-1iRGpr_YepvPkW7HD%3DEW%3Db%2B4w1m-YxS4fO%3D61o9O9Q%40mail.gmail.com.