Disentangling regional impacts of joint teleconnections using causal representation learning

TL;DR

DAG-VAE is introduced, a causal representation learning approach that embeds a physics-informed directed acyclic graph in the latent space of a variational autoencoder and jointly learns nonlinear reduced representations of large-scale modes of variability and their causal interactions.

Abstract

Understanding teleconnections of large-scale modes of climate variability is relevant for seasonal predictability and support a dynamical understanding of climatic changes. While numerical model experiments are the most common approach for investigating counterfactual climate responses, their conclusions are subject to model biases. Data-driven approaches offer a complementary perspective. Deep learning can extract reduced-dimensional patterns but usually lacks causal interpretability, while causal methods can disentangle signals in the presence of confounding yet are typically based on simple indices. Treating dimensionality reduction and causal inference separately thereby risks losing the teleconnection signal of interest. This paper introduces DAG-VAE, a causal representation learning approach that embeds a physics-informed directed acyclic graph in the latent space of a variational autoencoder. Combining deep learning with causal inference, the method jointly learns nonlinear reduced representations of large-scale modes of variability and their causal interactions. We apply DAG-VAE to disentangle the influences of the Pacific and Indian Oceans on the short rains over the Greater Horn of Africa. Trained on seasonal hindcasts, the method identifies dynamically meaningful representations and recovers spatial response patterns consistent with SST-replacement experiments. Trained on reanalysis data, DAG-VAE identifies a different response pattern to direct influence of the tropical Pacific, highlighting potential model biases and the value of DAG-VAE as a complementary, data-driven approach for estimating spatial causal response patterns from observations. Finally, we demonstrate the ability of the method to generate data-driven counterfactuals of extreme short rain seasons, with potential applications for forecast-based early action and scenario planning.

Figures (10)

• Figure 1: a) Illustration of the selected regions and variables used to study Greater Horn of Africa (GHA) precipitation, Indian Ocean (IO) and Pacific Ocean sea surface temperatures (SSTs). (b)Anomalies of precipitation over the GHA, SST patterns associated with a positive Indian Ocean Dipole and an El Niño season, showing October to December averages of the seasonal hindcast initialised in September 2015. The dashed black boxes indicate the regions used to compute the Dipole Mode Index and the Niño3.4 index.
• Figure 2: Illustration of the DAG-VAE method and architecture. $z_{TP}$, $z_{IO}$ and $z_{PR}$ refer to the low-dimensional representations identified by the model for the three high-dimensional input spaces: tropical Pacific SSTs, Indian Ocean SSTs and GHA precipitation. The red arrows refer to encoder and decoder neural networks that dimensionality-reduce the high-dimensional input space. The blue arrows refer to the directed acyclic graph (DAG) hypothesized in the latent space.
• Figure 3: Illustration of the learned causal representations in SEAS5, shown through reconstructed samples of the latent dimensions $z_{TP}$, $z_{IO}$ and $z_{PR1,2}$. The samples are created by first selecting equidistant samples from the means of the latent Gaussians, ranging from the minimum to the maximum value encountered in test set. Second, these samples are passed back through the decoder to reconstruct their original dimension.
• Figure 4: Intervention experiments conducted using DAG-VAE trained on SEAS5. Stippling indicates anomalies whose sign is consistent across 25 training runs, sampling the uncertainty introduced through the initial weights. a) The latent variable representing SSTs in the Indian Ocean, $z_{IO}$, is set to climatological (neutral) conditions, and samples are drawn from the latent variable representing SSTs in the Pacific, $z_{TP}$, shown in the top row after decoding. The effect of this intervention on precipitation over the GHA is estimated using the causal model fit in the latent space and shown in the bottom row, aligned to the corresponding sample s of $z_{TP}$. b) Here, $z_{TP}$ is set to neutral conditions and samples are drawn from $z_{IO}$ (top row), with the corresponding effect on precipitation over the GHA shown in the bottom row.
• Figure 5: Same as Figure \ref{['seas5_representations']}, but for ERA5.