This repository contains code for generating datasets of parcel model runs using PySDM, and using them to train machine learning emulators of aerosol activation, using approaches based on the work of Silva et al.

To generate a dataset on Caltech's HPC cluster, run the following:

sbatch generate_parcel_data.sh [dataset_name] [num_simulations] [num_modes]

And then to see the output logs displayed,

tail -f slurm.out

For instance,

sbatch generate_parcel_data.sh my_2modal_dataset 20000 2

will perform 20000 runs of simulations with 2-modal aerosol populations, creating the files datasets/my_2modal_dataset_train.csv, datasets/my_2modal_dataset_test.csv, and datasets/my_2modal_dataset_fail.csv - which are the generated train dataset, test dataset, and initial conditions of the runs that failed (due to a condensation solver failure or timeout).