Overview

RTnn (Radiative Transfer Neural Networks) is a PyTorch-based framework for training neural networks to model radiative transfer processes in climate science, particularly for Land Surface Models (LSM).

What is RTnn?

RTnn provides a flexible and efficient framework for:

  • Emulating radiative transfer: Replace expensive physical RT models with fast neural networks

  • Data preprocessing: Handle large climate datasets with multiple dimensions

  • Multiple architectures: Support for LSTM, GRU, Transformer, and FCN models

  • GPU acceleration: Leverage CUDA for fast training and inference

  • Comprehensive evaluation: Built-in metrics and visualization tools

Problem Statement

Radiative transfer calculations are among the most computationally expensive components in climate models. These calculations determine how solar radiation interacts with:

  • Vegetation canopies (absorption, reflection, transmission)

  • Atmospheric layers (scattering, absorption, emission)

  • Surface properties (albedo, emissivity)

Traditional physical models require solving complex radiative transfer equations (e.g., two-stream approximation, discrete ordinates method) at each grid point and time step, making them a significant computational bottleneck.

Solution Approach

RTnn addresses this challenge by training neural networks to learn the input-output mapping of radiative transfer processes:

Input variables:

  • Solar zenith angle (coszang)

  • Leaf area index (LAI) - collimated and isotropic

  • Leaf single scattering albedo (SSA) and phase function asymmetry (PSD)

  • Surface reflectance (rs_surface_emu)

Output variables:

  • Collimated and isotropic albedo

  • Collimated and isotropic transmittance

  • Absorption rates (channels 1-2 and 3-4)

Architecture Overview

Input Data (NetCDF) → DataPreprocessor → DataLoader → Model → Output
     ↓                      ↓              ↓          ↓         ↓
rtnetcdf_XXX_YYYY.nc   Normalization   Batching    Forward   Predictions
                      Variable groups   Shuffle     Pass     Unnormalized
                     Spatial/temporal                          Results
                          batching

Key Components

  1. DataPreprocessor: Handles loading and preprocessing of NetCDF files with multi-year and multi-processor data

  2. Model Architectures: Multiple neural network options including LSTM, GRU, Transformer, and FCN

  3. Evaluation Framework: Comprehensive metrics and loss functions for radiative transfer applications

  4. Command Line Interface: Easy training and inference without coding

Performance Highlights

  • Speed: Up to YYYx faster than physical RT models

  • Accuracy: R² > 0.95 across all output variables

  • Scalability: Efficient on multi-GPU and distributed systems

  • Data efficiency: Trained on ~5 years of data, validated on independent years

Next Steps