Knowledge-guided machine learning for disentangling Pacific sea surface temperature variability across timescales

TL;DR

This paper introduces a knowledge-guided autoencoder that disentangles overlapping Pacific SST variability modes across timescales, revealing their physical characteristics and interactions, and improving climate model analysis.

Contribution

The study develops a spectral-constrained autoencoder that identifies physically interpretable SST variability modes without predefined filters, advancing understanding of their interactions and model biases.

Findings

Separates ENSO-like and decadal modes with distinct spatial patterns

Shows decadal mode modulates ENSO diversity and timing

Reveals model-specific biases in ENSO representation

Abstract

Global weather patterns and regimes are heavily influenced by the dominant modes of Pacific sea surface temperature (SST) variability, including the El Ni\~no-Southern Oscillation (ENSO), Tropical Pacific Decadal Variability (TPDV), North Pacific Meridional Mode (NPMM), and the Pacific Decadal Oscillation (PDO). However, separating these modes of variability remains challenging due to their spatial overlap and possible nonlinear coupling, which violates the assumptions of traditional linear methods. We develop a Knowledge-Guided AutoEncoder (KGAE) that uses spectral constraints to identify physically interpretable modes, without the need for predefined temporal filters or thresholds. The KGAE separates ENSO-like modes on 2- and 3-7-year timescales and a decadal mode with characteristics reminiscent of the PDO and the NPMM, each with distinct spatial patterns. We demonstrate that the decadal mode modulates ENSO diversity (central Pacific versus eastern Pacific), and that a quasibiennial mode leads and follows the interannual mode, suggesting a role in ENSO onset and decay. When applied to climate model output, KGAEs reveal model-specific biases in ENSO diversity and seasonal timing. Finally, residual training isolates a primarily equatorial decadal mode, which may be a component of TPDV-related decadal variability, likely originating from advection linked to upwelling near the Gal\'apagos Islands and the South Equatorial Current. Our results highlight how machine learning can uncover physically meaningful modes of Earth system variability and improve the representation and evaluation of variability across models and timescales.