24AI Temporal Causal Analysis of Aerosol-Cloud Interactions in Atmospheric Models Using Time-Series Causal Neural Networks

The field of time series causal discovery has garnered significant interest across multiple disciplines. Nevertheless, many current methodologies operate under the assumption of stationarity, a premise that often fails in the context of aerosol-cloud interactions, where aerosols, clouds and precipitation change rapidly. To overcome this limitation, we present a new causal discovery framework, the Time-Series Causal Neural Network (TS-CausalNN), which utilizes convolutional neural networks. This innovative deep learning model is adept at uncovering complex relationships within time series data that are non-stationary, non-linear, and noisy. In particular, it generates directed acyclic graphs (DAGs) to represent independence relationships in temporal datasets, allowing for the simultaneous identification of both contemporaneous and lagged causal relationships.

Using the above proposed TS-CausalNN approach, we focus on a selection of warm marine boundary layer cloud cases over the northeastern Atlantic, employing large-domain Weather Research and Forecast (WRF) large-eddy simulations, single-column model (SCM) simulations with the DOE Energy Exascale Earth System Model (E3SM), and doubly periodic mode simulations of the Simplified Cloud Resolving E3SM Atmosphere Model (SCREAM). We prepare and standardize time series data for cloud, precipitation, and boundary layer state variables associated with the aerosol indirect effect to conduct temporal causal analyses of aerosol-cloud interactions. The causal links identified among the modeled processes will enhance our understanding of the physical processes involved. Additionally, comparing causal linkages from large-eddy simulations and E3SM results will help identify limitations in E3SM model parameterizations.

Acknowledgment: This work is supported by the DOE Office of Science Early Career Research Program and the ASR Program. This work was performed under the auspices of the U.S. DOE by LLNL under contract DE-AC52-07NA27344. LLNL-ABS-868598