Climate

DLESyM-Ocean: A Deep Learning Probabilistic Global Model for Simulating Present-Day Upper Ocean and Sea Ice

Zachary I. Espinosa, Nathaniel Cresswell-Clay, William Yik, Cecilia M. Bitz, Edward Blanchard-Wrigglesworth, Peter Harrington, David Pruitt, Michael S. Pritchard, Dale R. Durran

Featured August 16, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

This model uses AI to predict how oceans and sea ice will change over many years, creating multiple possible futures by adding randomness and using a special scoring system that makes sure its predictions look realistic across space.

In depth

The paper introduces DL ESy M-Ocean, a deep learning model for simulating global upper ocean and sea ice conditions. It generates probabilistic ensemble forecasts by injecting noise through conditional layer normalizations. A key innovation is the Almost Fair Patch Energy Score (afPES) loss function, which ensures spatial coherence in predictions and addresses degeneracies found in standard fair scoring rules, leading to stable multi-year simulations of complex Earth system dynamics.

Key Takeaways

1
The model generates probabilistic ensemble forecasts for upper ocean and sea ice, capturing a range of plausible future states.
2
It employs a novel Almost Fair Patch Energy Score (afPES) loss to ensure spatial coherence and stable training for multivariate predictions.
3
DL ESy M-Ocean demonstrates multi-year stability and accurately reproduces climatology, variability, and extreme events like marine heatwaves and El Niño.

Conceptual Flow

HIGH LEVEL

Generating Realistic Ocean & Ice Futures

The model takes current and past ocean data, adds a bit of controlled randomness, and then predicts future ocean and ice conditions, learning from its mistakes across small areas.

Past Ocean Data

Current Ocean Data

Atmosphere Info

Random Noise

Predict Future States

Multiple Future Scenarios

Accurate Multi-Year Climate Simulation

The model successfully predicts long-term ocean and sea ice patterns, including big events like heatwaves and El Niño, showing it understands how the ocean works.

Model Predictions

Real World Data

Compare & Verify

Matches Real Patterns

Captures Big Events

Stable Over Years

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.