Enterprise AI Analysis: AERIS: Argonne Earth Systems Model for Reliable and Skillful Predictions
Revolutionizing Earth System Prediction with Exascale AI
The paper introduces AERIS, a groundbreaking Earth Systems Model that leverages a 1.3 to 80B parameter pixel-level Swin diffusion transformer. It achieves 10.21 ExaFLOPS sustained mixed-precision performance on 10,080 Aurora nodes, marking the highest throughput in AI for Science to date. AERIS employs a novel parallelism strategy, SWiPe, to efficiently shard window-based transformers, enabling stable and scalable training at high resolutions. It outperforms the IFS ENS in medium-range forecasts and maintains stability over seasonal scales (up to 90 days), demonstrating the potential of billion-parameter diffusion models for advanced weather and climate prediction. The model showcases significant advancements in computational performance, scaling efficiency, and forecast skill for extreme weather events like hurricanes and heatwaves.
Unlocking Unprecedented Performance
AERIS redefines the capabilities of AI in scientific computing, achieving breakthrough performance and efficiency on the world's most powerful supercomputers.
0
Sustained Mixed-Precision Performance
0
Peak Mixed-Precision Performance
0
Weak Scaling Efficiency
0
Strong Scaling Efficiency
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
AERIS demonstrates unprecedented computational throughput and scaling efficiency on exascale supercomputers, enabling the training of billion-parameter models for high-resolution weather and climate prediction.
0
Sustained Mixed-Precision Performance on Aurora
| Feature | Traditional Parallelism | SWiPe (AERIS) |
|---|---|---|
| Feature |
|
|
| Communication Overhead |
|
|
| Activation Memory Usage |
|
|
| Scaling Efficiency |
|
|
Enterprise Process Flow
Input Image Partitioned
→
Windows Distributed Across Nodes
→
Windows Sharded Across GPUs (Ulysses SP)
→
Pipeline Parallelism Across Layers
→
Communication Merged/Overlapped
→
Efficient Parallel Computation
AERIS achieves competitive to superior forecast skill compared to state-of-the-art models like IFS ENS and GenCast, demonstrating unique stability across medium-range and seasonal scales.
0
Stable Forecasts on Seasonal Scales
Hurricane Laura Prediction (2020)
AERIS accurately predicted Hurricane Laura's track and rapid intensification up to 7 days before landfall, showcasing superior skill in forecasting extreme events compared to global numerical models. This demonstrates the model's ability to provide crucial lead time for disaster preparedness.
- ✓ Predicted track with minimal errors 7 days prior.
- ✓ Accurately forecasted rapid intensification.
- ✓ Enabled critical lead time for emergency response.
London Heatwave Prediction (2020)
The model successfully identified the intense August 2020 London heatwave more than a week in advance. All ensemble members captured the sharp temperature rise and return to climatology, closely matching ERA5 data. This highlights AERIS's reliability in predicting high-impact extreme weather events.
- ✓ Identified heatwave >1 week in advance.
- ✓ Ensemble members matched ERA5 temperature trends.
- ✓ Proved reliable for high-impact extreme events.
AERIS leverages a pixel-level Swin diffusion transformer with architectural and parallelism innovations to achieve stability, scalability, and long-range forecast skill.
0
Maximum Parameter Count
| Feature | Previous Data-Driven Models | AERIS |
|---|---|---|
| Transformer Type |
|
|
| Resolution & Patch Size |
|
|
| Generative Approach |
|
|
| Seasonal Stability |
|
|
Enterprise Process Flow
Initial Condition Input
→
Iterative Diffusion Steps (Ensembles)
→
Autoregressive Steps
→
Stable Forecasts to 90 Days
Calculate Your Potential ROI with AERIS
See how integrating AERIS into your operations can translate into significant efficiency gains and cost savings for your enterprise.
Strategic Implementation Roadmap
Our phased approach ensures a seamless integration of AERIS capabilities into your existing infrastructure, maximizing impact with minimal disruption.
Enhance Ensemble Spread
Implement initial condition perturbations and fine-tune stochastic churning schedule under TrigFlow to improve the diversity and spread of ensemble members without compromising skill. This directly addresses the current under-dispersive nature of AERIS ensembles.
Optimize Inference Cost with Consistency Distillation
Leverage consistency distillation to compress the model size and reduce inference steps to a single-step, significantly lowering the computational cost for new forecasts. Explore multi-step finetuning to further improve forecast skill with distilled models.
Improve Pipeline Parallelism Efficiency
Reduce pipeline bubble size by adopting zero-bubble pipeline parallelism strategies (e.g., 1F0B) to minimize GPU idle time during inter-stage data transfer. This will enhance overall throughput and scaling efficiency on large supercomputers.
Integrate Physical Priors and Data Assimilation
Develop methods to guide the diffusion process with physical priors and assimilate real-time observations to improve forecast accuracy and physical consistency. This addresses potential non-physical artifacts inherent in purely data-driven models.
Expand to Alternative Datasets and Longer Time Horizons
Explore training AERIS on higher-resolution inputs and alternative Earth system datasets (e.g., coupled ocean-atmosphere models) to extend forecast skill to even longer time horizons and broader applications.
Ready to Redefine Your Predictions?
Connect with our experts to explore how AERIS can transform your organization's approach to Earth system modeling and forecasting.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat