AI Climate Model Simulates 1000 Years in a Day, Outperforming Supercomputers and Slashing Energy Use

AI Model Simulates 1000 Years of Climate in Just One Day

Extreme weather events once considered “100-year” occurrences are now happening with alarming frequency. These events—floods, storms, wildfires—often break records for intensity and damage. By definition, a 100-year event has only a 1% chance of occurring in any given year. But determining whether such events align with current climate trends or are statistical outliers remains a challenge.

A graduate student named Nathaniel Cresswell-Clay has developed an AI model that simulates up to 1,000 years of today's climate using far less computing power than traditional methods. This model effectively captures atmospheric conditions like the low-pressure system over the central United States, which plays a major role in weather patterns.

Limitations of Traditional Forecasting

Standard weather and climate models rely on energy-intensive supercomputers housed at major research centers. While AI has made strides in weather forecasting over the past five years, most AI-driven models only provide accurate predictions up to 10 days ahead. However, longer-range forecasts are crucial for climate science and for preparing communities for future seasonal impacts.

AI-Powered Climate Simulation Breakthrough

In a study published in AGU Advances, researchers at the University of Washington introduced an AI model capable of simulating Earth's current climate and interannual variability for up to 1,000 years. Remarkably, it runs on a single processor and completes a forecast in just 12 hours. By contrast, the same simulation on a top-tier supercomputer would take about 90 days.

This model, called Deep Learning Earth System Model (DL ESy M), combines two neural networks—one representing the atmosphere and the other the ocean. While traditional Earth system models often link atmospheric and oceanic components, this is the first time AI-only models have done the same.

Overcoming Data Limitations

Training AI models requires large datasets. However, the most reliable daily weather data only extends back to around 1979, offering limited seasonal cycles for training seasonal forecasts. Despite this, DL ESy M was trained for one-day forecasts but still learned to capture seasonal variability.

The oceanic model in DL ESy M updates predictions every four days, reflecting the slower changes in sea surface temperature, while the atmospheric model updates every 12 hours. Plans are underway to add a land-surface model, which would help simulate interactions between soil, vegetation, and the atmosphere—complex relationships that are difficult to model with traditional equations but can be learned directly by AI.

Performance Compared to Leading Climate Models

DL ESy M was tested against four leading models from the sixth phase of the Coupled Model Intercomparison Project (CMIP6), which inform the reports of the Intergovernmental Panel on Climate Change (IPCC). The AI model performed better at simulating tropical cyclones and the seasonal cycle of the Indian summer monsoon.

In mid-latitude regions, DL ESy M captured month-to-month and year-to-year variability as well as CMIP6 models. It also effectively simulated atmospheric “blocking” events—patterns where atmospheric ridges cause prolonged hot, dry weather in some regions and cold, wet conditions in others. Many existing climate models struggle with blocking events, but DL ESy M matched the performance of physics-based models.

Implications for Climate Research

• DL ESy M uses significantly less computing power, reducing the carbon footprint of climate simulations.
• Its accessibility means researchers without supercomputer access can run complex climate experiments.
• The approach opens up possibilities to include more Earth system components, improving model accuracy over time.

This advancement offers a practical tool for assessing whether extreme weather events are within the expected range of natural climate variability or signs of changing climate patterns. The model’s efficiency and accuracy make it a valuable resource for climate researchers and policymakers alike.

For those interested in deepening their AI skills, especially in scientific applications, exploring specialized AI courses can provide valuable expertise. Check out Complete AI Training’s latest courses for more information.