This module uses recurrent neural networks (RNN) to predict a sequence of macroscopic observables (i.e. stress) of a granular material undergoing a given input sequence of strains. We have tested this module for training on DEM simulations using YADE. In the dataset we consider triaxial compressions of samples with different contact parameters.
[here could go the illustration of the kind of time series]
During the installation of grainLearning activate extra rnn
:
poetry install --extras "rnn"
or pip install .[rnn]
Note: This will install a version of tensorflow depending on your system. If gpu dependencies are not installed or activated, it will default to CPU version. Check the installation of tensorflow and/or re-install it following specific instructions for your hardware.
For MacOS with arm64 processor, we recommend to install tensorflow following this and install grainLearning without
rnn
extra. In this case if you want to have weights and biases you can install it viapip install wandb
.
There are three main usages of RNN module:
- Train a RNN with your own data.
- Make a prediction with a pre-trained model.
- Use a trained RNN in grainLearning calibration process.
For more details about how are these model built check the documentation.