Spatiotemporal Machine Learning Approaches for Atmospheric Composition Emulation in NASA GISS ModelE

Abstract Earth System Models (ESMs) rely on parameterizations to represent sub‐grid scale processes that cannot be explicitly resolved at typical model resolutions. However, maintaining full coupling between these parameterizations and other model components creates substantial computational demands. This challenge is particularly acute for atmospheric composition modules, where numerous aerosol species and constituents must be advected and processed at each model timestep. The resulting computational overhead severely constrains the feasibility of conducting long‐term, high‐resolution climate projections. To address these computational limitations, the NASA GISS‐E3 (ModelE) employs a Non‐Interactive Tracer (NINT) methodology, utilizing pre‐calculated monthly climatologies of atmospheric composition fields. While computationally efficient, this approach eliminates the dynamic feedbacks between meteorological variability and chemical processes. This work introduces a machine learning (ML) framework designed to bridge this gap by creating a “Smart‐NINT” system. Our approach leverages neural networks to approximate the advection, and removal terms that govern tracer evolution, allowing for meteorologically responsive composition fields without explicit tracer transport calculations. The methodology focuses on surface‐level dynamics, specifically targeting Black Carbon aerosols from wildfire sources. We trained our models using two years of fully coupled ModelE output (1950–1951). The models were evaluated for their ability to capture both temporal and spatial dependencies. Results revealed consistent performance across approaches, with averaging 0.80. The Convolutional Long Short‐Term Memory (ConvLSTM) architecture, which combines convolutional and recurrent neural networks for spatiotemporal processing, demonstrated the best performance.