Project Journal

1. Data Preprocessing - Completed, Tested. Done for both Ronak [11/25/2024] and Abhisek[12/04/2024]
2. Abhisek working on Case Studies 1 and 2 with DeepOnet [12/05/2024]
3. Ronak working on Case Studies 3 and 4. Made code changes to DeepONet to time march the model prediction process. Testing in progress. [12/05/2024]

Models to consider

1. DeepONet - Can use it for rectangular domains. Starting with this model.
2. FNO - Find a way to handle rectangular domains.
3. Poseidon/scOT - Can't handle rectangular domains. Checked with the authors. They want us to do it.
4. CNO - Not sure about this one.

Part 0 - Comment Request from Ethan

0. The objective for the changes:

   • We want to time march the model prediction process so that it can reuse previous predictions to make future predictions. e.g.,
     • Train x1, x2, x3, x4, x5 -> x6
     • Train x2, x3, x4, x5, x6* -> x7
     • Train x3, x4, x5, x6*, x7* -> x8
     • Train x4, x5, x6*, x7*, x8* -> x9
     • Train x5, x6*, x7*, x8*, x9* -> x10

   Here x6*, x7*, x8*, x9* are the past predictions made by the model. The model should be able to reuse these predictions to make future predictions.

1. Review Required on two files:

   ./model_training/models/deeponet.py
   ./model_training/data/dataset_fpo.py
   

2. Changes Made in deeponet.py

   • Added method on_train_epoch_end(). This allows us to perform the following operations at the end of each epoch:
     • Save the current prediction to a .npy file
     • Reload the Dataset object by discarding the first prediction and adding the new prediction
     • Do it every 100 epochs
     • Calls an update_data() method in the DataModule object to update the dataset

3. Changes Made in dataset_fpo.py

   • Added class FPODataModule and in it a method update_data(). This allows us to perform the following operations:

     • Load the current prediction from the .npy file
     • Performs the full range of time stepping calculations to update the dataset
     • Uses the time calculations to call the DataLoader object to update the dataset.

   • Added a new FPODatasetMix dataloader that allows us to mix the current prediction with the actual data to train the model, given a set of time indices.

Part - I : Data Preprocessing - Status : Completed, Tested. Done

0. Request an Interactive Node and activate the virtual environment on Nova:

   salloc -N 2 -n 72 -t 08:00:00 --mem 369G
   source /work/mech-ai/rtali/mlpythonenv/bin/activate
   

1. Check the data settings here:

   ./dataset_generation/data.ini
   

2. Run the following command to generate the dataset:

   python3 ./dataset_generation/data_prep.py
   

3. Two numpy tensor files will be generated:

   ./dataset_generation/fpo/in_data.npz
   ./dataset_generation/fpo/out_data.npz
   

Part - II : Model Training - Status : Initial Code Changes Done. Finalize after Code Review

0. Request an Interactive GPU Node and activate the virtual environment on Nova:

   salloc -N 1 -n 8 --gres gpu:a100:1 -t 8:00:00 --mem 369G
   source /work/mech-ai/rtali/mlpythonenv/bin/activate