Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bridging Idealized and Operational Models: An Explainable AI Framework for Earth System Emulators

Pouria Behnoudfar, Charlotte Moser, Marc Bocquet, Sibo Cheng, Nan Chen

TL;DR

This work tackles persistent biases in high-resolution Earth-system simulations by bridging idealized and operational representations through an explainable AI framework that is physically grounded. It introduces a physically augmented latent space, a data-driven short-term forecast, and a reconfigured data assimilation step that ingests pseudo-observations from idealized models to correct global fields while preserving interpretability. The approach yields substantial improvements in ENSO spatial structure, indices, statistics, and diversity within CMIP6-era CESM2, and demonstrates scalable, uncertainty‑quantified digital-twin capabilities. By enabling efficient multi-model integration and encouraging cross-community collaboration, the framework offers a generalizable path for enhancing complex dynamical systems beyond Earth-system applications, with explicit physically meaningful justification for each correction.

Abstract

Computer models are indispensable tools for understanding the Earth system. While high-resolution operational models have achieved many successes, they exhibit persistent biases, particularly in simulating extreme events and statistical distributions. In contrast, coarse-grained idealized models isolate fundamental processes and can be precisely calibrated to excel in characterizing specific dynamical and statistical features. However, different models remain siloed by disciplinary boundaries. By leveraging the complementary strengths of models of varying complexity, we develop an explainable AI framework for Earth system emulators. It bridges the model hierarchy through a reconfigured latent data assimilation technique, uniquely suited to exploit the sparse output from the idealized models. The resulting bridging model inherits the high resolution and comprehensive variables of operational models while achieving global accuracy enhancements through targeted improvements from idealized models. Crucially, the mechanism of AI provides a clear rationale for these advancements, moving beyond black-box correction to physically insightful understanding in a computationally efficient framework that enables effective physics-assisted digital twins and uncertainty quantification. We demonstrate its power by significantly correcting biases in CMIP6 simulations of El Niño spatiotemporal patterns, leveraging statistically accurate idealized models. This work also highlights the importance of pushing idealized model development and advancing communication between modeling communities.

Bridging Idealized and Operational Models: An Explainable AI Framework for Earth System Emulators

TL;DR

This work tackles persistent biases in high-resolution Earth-system simulations by bridging idealized and operational representations through an explainable AI framework that is physically grounded. It introduces a physically augmented latent space, a data-driven short-term forecast, and a reconfigured data assimilation step that ingests pseudo-observations from idealized models to correct global fields while preserving interpretability. The approach yields substantial improvements in ENSO spatial structure, indices, statistics, and diversity within CMIP6-era CESM2, and demonstrates scalable, uncertainty‑quantified digital-twin capabilities. By enabling efficient multi-model integration and encouraging cross-community collaboration, the framework offers a generalizable path for enhancing complex dynamical systems beyond Earth-system applications, with explicit physically meaningful justification for each correction.

Abstract

Computer models are indispensable tools for understanding the Earth system. While high-resolution operational models have achieved many successes, they exhibit persistent biases, particularly in simulating extreme events and statistical distributions. In contrast, coarse-grained idealized models isolate fundamental processes and can be precisely calibrated to excel in characterizing specific dynamical and statistical features. However, different models remain siloed by disciplinary boundaries. By leveraging the complementary strengths of models of varying complexity, we develop an explainable AI framework for Earth system emulators. It bridges the model hierarchy through a reconfigured latent data assimilation technique, uniquely suited to exploit the sparse output from the idealized models. The resulting bridging model inherits the high resolution and comprehensive variables of operational models while achieving global accuracy enhancements through targeted improvements from idealized models. Crucially, the mechanism of AI provides a clear rationale for these advancements, moving beyond black-box correction to physically insightful understanding in a computationally efficient framework that enables effective physics-assisted digital twins and uncertainty quantification. We demonstrate its power by significantly correcting biases in CMIP6 simulations of El Niño spatiotemporal patterns, leveraging statistically accurate idealized models. This work also highlights the importance of pushing idealized model development and advancing communication between modeling communities.
Paper Structure (35 sections, 13 equations, 20 figures)

This paper contains 35 sections, 13 equations, 20 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Latent space representation and physical augmentation
  4. Data-driven short-term model forecasts in the augmented latent space
  5. Bridging via assimilation of idealized model output
  6. Correcting structural bias via curriculum learning
  7. Results
  8. Setup
  9. Reanalysis data
  10. Definitions of Niño regimes and different types of the ENSO events
  11. Operational models
  12. Idealized models
  13. Machine learning configuration and training
  14. Data assimilation setup
  15. ENSO bias correction in the bridging model
  16. ...and 20 more sections

Figures (20)

  • Figure 2.1: Overview of developing the bridging model from XAI.
  • Figure 2.2: Schematic of the curriculum learning strategy for training the autoencoder and forecast models. The training data is gradually transitioned from purely operational model (OM) outputs to a blend of OM and reanalysis (REA) data.
  • Figure 3.1: Comparison of different variability patterns, measured by the standard deviation, among CESM2, GODAS reanalysis, and the proposed bridging model. The left, middle, and right columns show results from CESM2, GODAS, and the bridging model, respectively. The top row displays sea surface temperature (SST, shading) and zonal wind stress ($\tau_x$, contours). The bottom row shows thermocline depth ($H$, shading) and subsurface temperature (TSUBA, contours).
  • Figure 3.2: Comparison of equatorial Pacific SST and Niño indices from the CESM2 model, GODAS reanalysis, and the bridging model. The left panel shows the SST along the equator (where the result using the idealized model is also included), while the right panel displays the Niño 3, Niño 3.4, and Niño 4 indices.
  • Figure 3.3: Statistical properties of the Niño 3, Niño 3.4, and Niño 4 indices for the CESM2 model, GODAS reanalysis, and the bridging model. The four rows present the probability density functions (PDFs), autocorrelation functions (ACFs), power spectra, and seasonal variation of SST. For the bridging model, the red shaded area represents the 95% confidence interval derived from 100 ensemble members, each with a length equal to the 40-year GODAS reanalysis period.
  • ...and 15 more figures