Physics-Guided Multi-Task Learning for Subgrid Scale Turbulence Parameterization: A Comparative Study of Physics Integration Strategies

The files contain ML model training and evaluation scripts, along with the associated results, evaluation metrics, taining plots wtih task weighting experiments and the trained ML model checkpoints. The scripts folder also contains a README file for instructions to use the srcipts. Due to space constraints, all the actual training and evaluation data sets could not be shared here. Hence, the repository contains the features and targets used in training the models, which were extracted from MONC simulation NETCDF files. The simulations are based on the Met Office–NERC Cloud Model (MONC), whose publicly available reference implementation is maintained at https://github.com/MetOffice/monc and https://github.com/EPCCed/monc. The experiments in this study were executed using a MONC branch obtained via the Met Office Science Repository Service (MOSRS), with local modifications limited to additional diagnostics and data extraction routines. Due to data governance constraints, the derived datasets are not publicly redistributed here; however, the MONC reference codebase, simulation configurations, and data‑extraction methodology are fully described in the paper, allowing independent reproduction of the datasets. For supporting the reproducibility of the simulations, this repository contains the MONC testcase configuration files required for reproducing the raw simulations for the Radiative Convective Equilibrium (RCE) and Atmospheric Radiation Measurement (ARM) simulations using the publicly availaible version of MONC. The users should remove the '.txt' extension from the '*.mcf.txt' files and place them in the respective 'testcases' path inside the MONC directory. Please refer to the official documentation for further details. If anyone is interested in using the NETCDF data sets from MONC or to learn more about reproducing the raw simulation data from MONC, they are encouraged to please contact the authors for more details and support regarding this, including data access and any further queries they might have about the paper. Please acknowledge and cite this resource, if used in any publication or other future works. Acknowledgements: This work was supported by the University of Reading Advancing the Frontiers of Earth System Prediction (AFESP) Programme [A3720300] and the UK Met Office. We acknowledge JASMIN for computational resources and technical support.