PCNN-GCM: Physics Constrained-Neural Network GCM

Introduction

Climate change is driving more frequent and intense extreme weather events, which have significant impacts on ecosystems and communities worldwide. General Circulation Models (GCMs) are essential tools for understanding and projecting climate change over various time scales. However, a major challenge for modern GCMs is accurately representing small-scale atmospheric processes—particularly cloud dynamics and convection—due to their relatively coarse spatial resolution (~50 km).

Traditional methods for addressing subgrid processes in GCMs include:

• Subgrid Parametrization Models: These use simplified empirical relationships or theoretical approximations to represent small-scale processes. Although computationally efficient, these models often introduce significant uncertainties into climate projections.
• Super Parametrization Models: By embedding high-resolution Cloud Resolving Models (CRMs) within each grid cell of a GCM, these models explicitly resolve small-scale dynamics, leading to improved accuracy. However, the high computational cost of this approach makes it impractical for long-term simulations.

Recently, hybrid modeling approaches that combine physical principles with machine learning have shown promise. By leveraging high-resolution data from CRMs, deep learning (DL) emulators can be developed to model small-scale processes more efficiently. Despite these advances, ensuring long-term simulation stability remains a critical challenge.

To address this, we introduce CondensNet—a novel neural network architecture that enforces a key physical constraint: water vapor saturation regulation. CondensNet is composed of two main components:

• BasicNet: Responsible for learning robust cloud representations.
• ConCorrNet: A condensation correction network that adaptively refines BasicNet's outputs to prevent water vapor oversaturation, ensuring both physical consistency and accuracy.

By integrating CondensNet with the Community Atmosphere Model (CAM), we construct the PCNN-GCM (Physics Constrained-Neural Network GCM) framework. This hybrid DL-GCM framework operates under real-world conditions (using prescribed sea surface temperatures, sea ice concentrations, and coupling with a land model), offering a significant step forward in achieving long-term, stable climate simulations with improved computational efficiency.

This repository contains two main modules:

• CondensNet/:
  The implementation of our novel CondensNet architecture, including its submodules (BasicNet for cloud representation and ConCorrNet for condensation correction) along with training scripts and configuration files.

• host-GCM/:
  A GCM framework based on NCAR CESM 1.1.1 that supports both deep learning parameterization (as implemented by CondensNet) and traditional super-parameterization approaches, forming the host component of the PCNN-GCM framework.

Content

This repository is organized into the following main directories:

• CondensNet/
  Contains the complete implementation of our novel CondensNet architecture, including:

  • Neural Network Design:
    The design of BasicNet for learning cloud representations and ConCorrNet for adaptive condensation correction.
  • Training Scripts:
    All scripts and configuration files needed to train CondensNet using high-resolution simulation data.
  • Inference Pipeline:
    Modules for deploying the trained network for model inference within a climate simulation framework.
  • Utilities and Documentation:
    Additional tools and documentation to help reproduce experiments and facilitate further research.

• host-GCM/
  Houses the experimental GCM framework based on NCAR CESM 1.1.1. This component is designed to support:

  • DL Parameterization Component:
    Supports calling CondensNet to solve the subgrid Parameterization process.
  • Super-Parameterization:
    An implementation inspired by the Superparameterized CAM (SPCAM) as detailed on the UCAR Wiki.
  • Coupling Mechanism:
    Seamless integration between the host GCM and CondensNet, enabling stable and efficient hybrid climate simulations.

• example.case/ The example.case directory adheres to the NCAR CESM standard directory structure, ensuring that all configuration, build, and run files are organized according to NCAR’s guidelines. For more details on the standard CESM case organization, please refer to the official CESM User Guide and NCAR CESM website:

  • NCAR CESM User Guide
  • NCAR CESM Official Website

Data

The dataset currently comprises training and test sets. Each file is stored in npz format and is named using the YYYY-MM-DD.npz convention. Within each file, two keys are provided:

• data_x: Neural network input data.
• data_y: Neural network output data.

Each npz file contains 48 SPCAM simulation time steps (with each time step representing 30 minutes), covering a full day of simulation. The dataset is accessible via CondensNet.