DNN_OSBLMixing

This repository is the computer codes to train and test a deep neural network (DNN) model for vertical mixing by ocean surface boundary layer turbulence described in Liang et al. [2022].

A. Download deep learning codes

Create a local copy of the codes by

git clone https://github.com/lsuocean/DNN_OSBLMixing/

There are two files under the directory

• train.py – script to train a DNN model

• test.py – script to run an offline 1D prediction for 9 days, using the trained DNN model.

The codes are written in Python and use Python packages "tensorflow" and "keras"

B. Create training and testing data

Create training/testing datasets using either observational data or numerical solutions. In Liang et al. [2022], we use multi-year large-eddy simulation (LES) solutions for the Ocean State Papa as training data (see Liang et al. [2022] for details). If you are interested in our LES solutions, feel free to send me an email to Dr. Jun-Hong Liang (jliang@lsu.edu).

The datasets include

• ocean_train.nc: the training dataset
• ocean_valid.nc: the validation dataset
• norm_paraz.npz: our DNN model predicts normalized temperature and salinity change rate over time. This file contains parameters to convert normalized outputs to the true values.

C. Train the DNN model

• Run the training codes

python3 codes/train.py

After the training is completed, two files are created:

1. cong_vs_epoch.dat: a record of the training and validation losses over epoch during DNN training process
2. weights_best.h5: the best trainable DNN parameters during the training process.

D. Test the DNN model

• Run the prediction codes

python3 codes/test.py

Steps C and D are repeated to select the optimal model hyperparameters.

---

Details of the model and its testing is in: Liang, J.H., Yuan, J., Wan, X., Liu, J., Liu, B., Jang, H. and Tyagi, M., 2022. Exploring the use of machine learning to parameterize vertical mixing in the ocean surface boundary layer. Ocean Modelling, p.102059. doi.org/10.1016/j.ocemod.2022.102059

If you have any questions regarding the model, please feel free to open an issue or send an email to Dr. Jun-Hong Liang (jliang@lsu.edu).