SIParCS 2026 - LJ Dunphy

Introducing CREDIT-chem: A Hybrid Physics-Machine Learning Framework for Modeling Atmospheric Aerosol Chemistry

Recorded Talk

The simulation of atmospheric chemistry and aerosols, such as black carbon, is extremely important for both climate simulations and modeling air quality. However, our current physical chemistry solvers are very computationally expensive, and coupling them with a meteorology or land model is even slower. To solve this, we propose a hybrid physics-based and machine learning (ML) approach to emulate the chemistry. Our model is trained on the Community Atmosphere Model (CAM) chemistry module, outputting black carbon aging, condensation, and deposition, and a semi-lagrangian advection scheme for physics-based transport. The model first applies the semi-lagrangian advection scheme to transport aerosol tracers, after which the ML model predicts the local column chemistry and microphysical processes. Unlike traditional global-field ML models that predict an entire global state in one pass, this columnar design allows columns to be evaluated independently and in parallel. Additionally, since each vertical column is independently predicted, we have significantly more training examples for generalization, and we can apply more explicit physical constraints. Preliminary 1-step rollout results show substantial speed ups in emulation, while retaining accuracy to CAM6-chem outputs. Airborne black carbon concentration model outputs show ~0.97-0.98 correlation coefficients to CAM6-chem targets, and ~0.83 for cloudborne black carbon. Level-wise correlation coefficients for airborne black carbon exceeded 0.999 in parts of the lower atmosphere as well. A 1-hour forecast at approximately 1-deg global resolution with our model runs in 2-4 seconds on 1 A100 40GB GPU, multiple orders of magnitude faster than CAM6-chem.

Mentors: David John "DJ" Gagne, Charlie Becker, Alma Roux, John Schreck

Slides and poster