RNNpaper

This repo contains code and data for the paper draft "Physically Interpretable Emulation of a Moist Convecting Atmosphere with a Recurrent Neural Network" by Qiyu Song and Zhiming Kuang. Following the steps below, one should be able to reproduce all results included in the paper.

Step 1: Generate Training Data

We use The System for Atmospheric Modeling (SAM) model to generate our data. The source code of the model with our modifications is in SAM_v6.11.7/SRC_noisywave/. We ran several groups of experiments with different configurations.

A. Random forcing experiments

For the 3 experiments to identify a linear model, we use an ensemble size of 1024. User should use prm.spinup_1024 as spinup configuration and these prm files 1024_4_1, 1024_4_2, 1024_4_3 as experimental configurations, and modify SAM_v6.11.7/SRC_noisywave/domain.f90 as:

    integer, parameter :: nx_gl = 1024 ! Number of grid points in X
    integer, parameter :: ny_gl = 1024 ! Number of grid points in Y
    integer, parameter :: nsubdomains_x  = 32 ! No of subdomains in x
    integer, parameter :: nsubdomains_y  = 32 ! No of subdomains in y

For the majority of experiments, we use an ensemble size of 256. Users can refer to prm files prm.spinup_256_0, prm.spinup_256_1 for spinup and SAM_v6.11.7/RCE_randmultsine/prm.run_256_msinefx* for experimental configurations. The domain setup should be:

    integer, parameter :: nx_gl = 512 ! Number of grid points in X
    integer, parameter :: ny_gl = 512 ! Number of grid points in Y
    integer, parameter :: nsubdomains_x  = 16 ! No of subdomains in x
    integer, parameter :: nsubdomains_y  = 16 ! No of subdomains in y

For all these and following SAM experiments, first compile the source code for an executable file and use the resub.ens file in the case directory to submit a job to a cluster (current version only reflect the setup on Harvard Cannon cluster).

B. Coupled wave forcing experiments

First, go to SAM_v6.11.7/RCE_noisywave/ and run the spinup experiment. Then use run_batch_noisywave.sh to generate case folders for different wavenumbers. We used two different values (1 and 2) in line 5 of that file, therefore having 2 experiments for each wavenumber differed by initial random seeds. Submit all experiments using submit_exps.sh.

After running the experiments, convert .stat files to .nc files using stat2nc, which can be compiled in SAM_v6.11.7/UTIL/. Then use RNN_train_test/extract_data.ipynb to extract the data for training.

Step 2: Train the Model

2.1 Identify a linear model

cd linear_model_paper
sbatch identification.run_4x

The identified linear model should be in linear_model_paper/model/. Load the model and write the variables A,B,C,K,NoisePattern in the sys variable to a .txt file, which will be used in next steps.

(optional) For a benchmark comparison, a linear response model without memory can be calculated as in linear_model_paper/get_linear_response_matrix.m, which will write the linear response matrix to another .txt file.

2.2 Train the RNN model

We perform a 2-stage training for the model. In each stage, first train for h0 and then for all parameters. Make modifications to model script (line 184-187) and training script (line 23-36) accordingly. Then submit the training by

cd RNN_train_test
sbatch train_ultimaternn_includewave_addp_lightning

The learned model will be a checkpoint file with the lowest validation error. The authors mostly used this model for following analysis.

2.3 Other training

Train a model without memory: model script, training script

Train a model with only the random forcing dataset: use the intermediate model during normal training

Train a model with only the coupled-wave forcing dataset: model script, training script, model

Step 3: Analysis

This part is included in several notebooks:

• offline tests: RNN, model without memory
• online tests: RNN, model without memory, fitting initial hidden state for RNN

---

This page is last updated on May 28, 2025. For any questions regarding the code or data, please contact Qiyu Song (qsong@g.harvard.edu).