RTnn

RTnn (Radiative Transfer Neural Networks) is a PyTorch-based framework for emulating radiative transfer processes in climate models, with a primary application to Land Surface Models (LSM). It provides neural network surrogates that replace computationally expensive physical radiative transfer schemes while preserving accuracy.

GitHub repository: https://github.com/kardaneh/RTNN

Key Features

  • Multiple architectures: LSTM, GRU, Transformer, FCN, MLP

  • Climate data support: Native NetCDF4 handling with multi-year, multi-process workflows

  • GPU acceleration: CUDA support with multi-GPU training

  • Comprehensive evaluation: Built-in metrics (NMAE, NMSE, R²) and diagnostics/visualization tools

  • Flexible preprocessing: Multiple normalization methods (minmax, standard, robust, log1p, sqrt)

  • Command-line interface: Training and inference directly from CLI without code changes

Applications

  • Emulation of canopy radiative transfer in vegetation models

  • Acceleration of climate model simulations

  • Data assimilation and uncertainty quantification

  • Sensitivity analysis of radiative transfer parameters

Performance

  • Up to YYY× speed-up compared to physical radiative transfer models

  • Minimal accuracy loss (typically R² > 0.9999)

  • Scalable to large datasets with multi-GPU training

Indices and tables