As extreme and unpredictable weather events increase globally, the ability to forecast conditions beyond two weeks—known as subseasonal forecasting—has become crucial. Accurate projections in this timeframe enable agriculture, energy providers, fisheries, and government agencies to make informed, proactive decisions that minimize losses and protect communities.
This advanced forecasting empowers farmers to plan crops and manage water resources in drought-prone zones. Utility companies can adjust for potential energy surges or shortages, while fisheries take steps to safeguard against sudden marine heatwaves. Importantly, governments can mobilize disaster response teams in advance, allocate resources for heat risk, or distribute aid before an event actually strikes.
Transforming Forecasting with AI and NVIDIA Earth-2
Artificial intelligence is transforming the field of climate and weather prediction. Over the past two years, the integration of advanced AI models into operational forecasting has accelerated, with platforms like NVIDIA Earth-2 at the forefront. This ecosystem provides scalable tools not just for scientists validating new models, but also for developers deploying and customizing large-scale predictive solutions.
By enabling efficient construction and deployment of powerful neural network models, Earth-2 is democratizing access to high-resolution, ensemble-based weather prediction. This unlocks more accurate results at a fraction of the computing cost associated with traditional, physics-based models.
Unlocking Large Ensemble Forecasts
A major breakthrough of AI-driven weather forecasting is the capability to run extensive ensemble predictions. Traditional methods are often constrained by hardware limits, but AI models built for Earth-2—such as FourCastNet V2 (SFNO) and Deep Learning Earth System Model (DLESyM)—can scale to produce forecasts from thousands of slightly different “initial states” or model weights.
This large ensemble strategy, including methods like Bred Vector/Multi Checkpoint (BVMC), captures a broad range of possible weather outcomes and quantifies uncertainty in the forecast. Insurance firms, researchers, and environmental agencies are leveraging such forecasts to better manage climate-related risks and design robust strategies.
Deep Learning Earth System Model (DLESyM): A Step Forward
DLESyM, supported by the Earth2Studio toolkit, stands out for its innovative architecture. This deep learning model combines separate AI modules for atmosphere and ocean, linked to dynamically track variations in sea surface temperatures and atmospheric patterns on a global scale.
Designed for efficiency and scalability, DLESyM uses a U-Net framework optimized for spherical data grids, sidestepping some of the typical limitations seen in conventional models. Its structure allows it to produce accurate forecasts up to several months ahead, maintaining realistic error margins and demonstrating remarkable stability under long, uninterrupted runs.
Below is an example workflow using Earth2Studio that showcases how straightforward it is to prepare data, configure model runs, and generate a 60-day subseasonal forecast:
python
package = DLESyMLatLon.load_default_package()
model = DLESyMLatLon.load_model(package).to(device)
data = ARCO()
io = KVBackend()
ic_date = np.datetime64("2021-06-15")
n_steps = 16
input_coords = model.input_coords()
output_coords = model.output_coords(input_coords)
inp_lead_time = input_coords["lead_time"]
out_lead_times = [
output_coords["lead_time"] + output_coords["lead_time"][-1] * i
for i in range(n_steps)
]
output_coords["lead_time"] = np.concatenate([inp_lead_time, *out_lead_times])
io = run.deterministic(
[ic_date], n_steps, model, data, io, output_coords=output_coords
)
Probabilistic Forecasting and Ensemble Strategies
Subseasonal-to-seasonal (S2S) weather events are naturally uncertain, and so forecasting approaches must reflect probabilistic outcomes rather than simple point predictions. AI-powered models quantify the likelihood that a season falls in the upper, middle, or lower tier relative to historical climate, focusing on variables like temperature and precipitation.
Earth2Studio now supports sophisticated ensemble generation with different model architectures, allowing experts to test new strategies for improving forecast skill, optimize ensemble size, or reduce compute and data storage costs via selective output saving.
For example, to generate and analyze ensemble predictions for significant events like the 2021 Pacific Northwest heatwave, a configuration such as the following is used:
text
nperturbed: 4
ncheckpoints: 16
batch_size: 4
defaults:
- forecast_model: dlesym
- perturbation: gaussian
This setup enables nuanced exploration into how initial-condition variation or alternate model checkpoints affect forecast uncertainty and reliability—critical information for sectors facing major financial or safety risks due to weather extremes.
Benchmarking Model Skill and Future Developments
Validating the accuracy and skill of long-range forecasts is essential for real-world application. Earth2Studio includes routines to assess predictions using established benchmarks, such as the European Centre for Medium-Range Weather Forecasts (ECMWF) IFS ENS model.
By integrating with community initiatives like the AI Weather Quest competition, NVIDIA and its collaborators are enabling researchers worldwide to compare models, quickly iterate on new architectures, and share progress through open source toolkits.
Efficient multi-GPU inference and scoring capabilities mean an entire year’s worth of ensemble forecasts can now be evaluated in mere hours—dramatically speeding up the pathway from model development to operational impact.
Visualizing and Scoring Forecasts
Using built-in Earth2Studio scoring tools, practitioners can rapidly assess model performance. For instance, z500 (geopotential height) score comparisons from 2018 show that DLESyM predictions are competitive with leading operational models for weeks three through five of a forecast—a crucial, previously hard-to-predict timeframe.
Alongside direct model validation, users can leverage AI Weather Quest routines within Earth2Studio to assign robust scores (like CRPS and RPSS) for various forecast variables and time ranges:
python
for var in aiwq_variables:
fcst_data, fcst_coords = load_forecast_for_aiwq(io_backend, ic, var)
rpss_wk3, rpss_wk4 = compute_aiwq_rpss(fcst_data, fcst_coords, var)
write_aiwq_scores(score_io_dict, rpss_wk3, rpss_wk4, var)
Conclusion: Transforming Climate Risk Management
The integration of AI and scalable cloud tools in NVIDIA Earth-2 is revolutionizing subseasonal weather forecasting. Sectors dependent on climate—agriculture, energy, infrastructure, logistics, and public policy—are positioned to benefit from forecasts that are more accurate, timely, and operationally useful than ever before.
With open-source access, robust model validation tools, and disciplined benchmarking against global standards, Earth-2 is setting a new bar for longer-term weather prediction. The ongoing development and community collaboration promise continued progress, driving safer and smarter decisions in the face of climate uncertainty.
Read more such articles from our Newsletter here.
