Interactive Atmospheric Composition Emulation for Next-Generation Earth System Models

TL;DR

The paper tackles the computational burden of interactive atmospheric composition in Earth System Models by introducing Smart NINT, a spatiotemporal ML emulator that predicts 3D BCB concentrations from surface emissions and meteorology. The approach uses dual data modalities (2D surface variables and 3D forcings), spatial encoders, a Spatial Feature Transform module, and a ConvLSTM backbone within a two‑tier training regime, focusing on the first 20 tropospheric levels of ModelE. Key findings show high fidelity near the surface (e.g., $R^2\approx0.92$, $r\approx0.96$ at Level 1) with performance gradually decreasing at higher levels, and reasonable generalization to periods beyond the training domain. The results demonstrate that architectural provisions for time and vertical coupling can yield physically plausible, long‑term predictions, enabling more affordable, high‑resolution climate projections while highlighting region‑specific biases during extreme wildfire years.

Abstract

Interactive composition simulations in Earth System Models (ESMs) are computationally expensive as they transport numerous gaseous and aerosol tracers at each timestep. This limits higher-resolution transient climate simulations with current computational resources. ESMs like NASA GISS-ModelE3 (ModelE) often use pre-computed monthly-averaged atmospheric composition concentrations (Non-Interactive Tracers or NINT) to reduce computational costs. While NINT significantly cuts computations, it fails to capture real-time feedback between aerosols and other climate processes by relying on pre-calculated fields. We extended the ModelE NINT version using machine learning (ML) to create Smart NINT, which emulates interactive emissions. Smart NINT interactively calculates concentrations using ML with surface emissions and meteorological data as inputs, avoiding full physics parameterizations. Our approach utilizes a spatiotemporal architecture that possesses a well-matched inductive bias to effectively capture the spatial and temporal dependencies in tracer evolution. Input data processed through the first 20 vertical levels (from the surface up to 656 hPa) using the ModelE OMA scheme. This vertical range covers nearly the entire BCB concentration distribution in the troposphere, where significant variation on short time horizons due to surface-level emissions is observed. Our evaluation shows excellent model performance with R-squared values of 0.92 and Pearson-r of 0.96 at the first pressure level. This high performance continues through level 15 (808.5 hPa), then gradually decreases as BCB concentrations drop significantly. The model maintains acceptable performance even when tested on data from entirely different periods outside the training domain, which is a crucial capability for climate modeling applications requiring reliable long-term projections.

Figures (12)

• Figure 1: Vertical profiles of BCB concentration across 20 atmospheric levels for four different years: (a) 1853, (b) 2012, (c) 2020, and (d) 2021.
• Figure 2: Annual mean BCB column burden. The first column is the ground truth from ModelE simulations. The second column is the machine learning predictions, and the third column represents the bias (prediction - ground truth).
• Figure 3: Vertical profiles of BCB concentration across the entire 62 atmospheric levels for four different years of the test set.
• Figure 4: Spatial distribution of annual average BCB concentration during 2012 across eight different pressure levels from 979 hPa (Level 1) to 171 hPa (Level 35). Note the different color scales for each level, with concentrations decreasing by multiple orders of magnitude at higher altitudes.
• Figure 5: Spatiotemporal ML model, equipped with spatial feature extraction and modality fusion.