Smart-NINT: Spatio-Temporal Machine Learning Approaches for Atmospheric Composition Emulation in NASA GISS ModelE

This repository presents Smart NINT, a spatiotemporal machine learning (ML) framework designed to emulate the computationally expensive process of interactive atmospheric composition transport within Earth System Models (ESMs), specifically tested using the NASA GISS ESM V3.0 (ModelE).

---

Installation

conda create -n niswan -c conda-forge python=3.12
conda activate niswan
conda install -c conda-forge xarray dask netCDF4 bottleneck
conda -c conda-forge scikit-learn seaborn xesmf
pip install torch torchvision lightning tqdm

Overview

ESMs often rely on prescribed monthly tracer concentration fields (Non-Interactive Tracers, NINT) to save computational resources, which sacrifices the ability to capture real-time aerosol-climate feedbacks. Smart NINT addresses this limitation by using a spatiotemporal ML architecture to interactively calculate tracer concentrations, such as Black Carbon from Biomass Burning (BCB), based on real-time surface emissions and meteorological data.

This approach transforms AI climate modeling from basic numerical solver mimicry to spatio-temporal prediction. By incorporating architectural components that capture trends, seasonal patterns, and cyclical behaviors, the model can deliver accurate long-term forecasts which is essential for multi-decadal to centennial climate projections. This enables higher-resolution, reliable climate simulations without the prohibitive computational costs of full physics parameterizations.

---

Key Features

• Computational Efficiency: Significantly reduces the cost of simulating composition transport, enabling higher-resolution and longer transient climate simulations.
• Interactive Feedback: Emulates interactive emissions, allowing the climate model to capture real-time feedback between aerosols and other climate processes.
• Spatiotemporal Architecture: Utilizes a ConvLSTM-based architecture with an inductive bias specifically designed to capture the complex spatial and temporal dependencies in tracer evolution.
• High Performance: Achieves excellent performance for BCB concentrations, with R² values of 0.92 and Pearson $r$ of 0.96 at the surface level, maintaining acceptable performance even outside the training domain.
• Focus on Vertical Dynamics: Incorporates a preprocessing module to effectively fuse 2D emission data with 3D meteorological forcings across the first 20 vertical levels (up to 656 hPa), where most short-term BCB variation occurs.

---

Model Setup (Conceptual)

The model replaces the full physics solver for tracer transport with an spatiotemporal ML component:

Figure 1. Concept of spatiotemporal ML-based modeling for predicting BCB concentration for climate modeling.

$$\text{Concentration}(t) = \text{Smart NINT} \left( \text{Emissions}(t), \text{Meteorology}(t), \text{Previous State} \right)$$

Inputs

• 2D Surface Emissions: Black Carbon from Biomass Burning (BCB).
• 3D Meteorological Data: Forcings from ModelE (e.g., pressure, wind, temperature).

Architecture Highlights

1. Preprocessing Module: Fuses 2D emission data and 3D forcings into a unified spatiotemporal representation.
2. Core Model: A spatiotemporal deep learning model (e.g., ConvLSTM) that learns the complex transport and mixing processes.

---

Future Work

We aim to extend this methodology to other aerosol and gaseous tracers and integrate the Smart NINT framework into the operational ModelE for comprehensive long-term climate projections.

---

Reference

Please cite the following papers when referencing this work or any of the foundational concepts discussed, and use the provided links to find the manuscripts:

• Manuscript at NeurIPS 2025

  @article{erfani2025interactive,
    title={Interactive Atmospheric Composition Emulation for Next-Generation Earth System Models},
    author={Erfani, Seyed Mohammad Hassan and Lamb, Kara and Bauer, Susanne and Tsigaridis, Kostas and van Lier-Walqui, Marcus and Schmidt, Gavin},
    journal={arXiv preprint arXiv:2510.10654},
    year={2025}
  }

• Manuscript at NeurIPS 2024

  @inproceedings{erfani2024spatiotemporal,
    title={Spatio-Temporal Machine Learning Models for Emulation of Global Atmospheric Composition},
    author={Erfani, Mohammad and Lamb, Kara and Bauer, Susanne and Tsigaridis, Kostas and van Lier-Walqui, Marcus and Schmidt, Gavin},
    booktitle={NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning},
    url={https://www.climatechange.ai/papers/neurips2024/66},
    year={2024}
  }