Dataset for training and evaluating LMP-Tr cloud microphysics parameterization

This dataset contains the training and testing data used for the LMP-Tr cloud microphysics parameterization model. The data were derived from WRF simulations of 30 precipitation events over Southern China during 2018–2020. All samples are model-generated outputs rather than observational measurements. Each precipitation event contains more than 270,000 samples, with more than 8 million samples in total.

The data are provided in MATLAB .mat format. Each .mat file corresponds to one precipitation event and contains two main variables, x and y. The variable x is the model input and is reshaped in the code as (-1, 50, 15), representing 50 vertical layers and 15 input variables for each sample. The variable y is the prediction target and is reshaped as (-1, 50, 12), representing changes in 12 cloud microphysical variables across 50 vertical layers. The files are organized according to the experimental setup used in the manuscript:

traindata/: data1–data20, used for training and validation;
testdata/: data21–data30, used as the independent test set.

The WRF simulations were conducted using WRF version 4.2.1 with the Thompson microphysics scheme. The simulation domain covers Southern China, with a horizontal resolution of 3 km, a grid size of 577 × 481, and 50 vertical layers. The domain extends from 104.8895°E to 122.4543°E and from 16.2589°N to 29.3032°N. Initial and boundary conditions were derived from European Centre for Medium-Range Weather Forecasts (ECMWF) operational analysis data.

The variables include mass mixing ratios of water vapor, cloud water, cloud ice, rainwater, snow, and graupel (QV, QC, QI, QR, QS, QG); number concentrations of cloud ice, rainwater, and cloud water (NI, NR, NC); water-friendly and ice-friendly aerosol concentrations (NWFA, NIFA); and environmental variables including temperature, pressure, vertical wind speed, and layer thickness (T, P, W, DZ). Some variables are used as both input and output variables, while the environmental variables are used as input variables only.

Official WRF v4.2.1 source code:
github.com

ECMWF data portal:
data.ecmwf.int

This Zenodo record archives the processed machine-learning dataset derived from WRF outputs, not the original ECMWF data products. The ECMWF data were used only to provide the initial and boundary conditions for the WRF simulations.