NASA and IBM Unveil Surya: First Foundation Model for Solar Physics

The Sun Just Got an AI Upgrade

For the first time in heliophysics history, researchers can leverage a foundational AI model purpose-built for solar analysis. NASA-IBM's Surya (Sanskrit for "Sun") represents a quantum leap in our ability to understand and predict solar behavior. Trained on 257TB of full-resolution (4096×4096 pixel) data from NASA's Solar Dynamics Observatory (SDO), this transformer-based model processes 13 channels across AIA and HMI instruments at 12-minute cadence—capturing the Sun's dynamics in unprecedented detail.

Architecture: Solar-Specific Innovations

Surya's design addresses unique solar data challenges through three core innovations:

1. Spectral Gating Blocks

   "Learnable frequency-domain filters adaptively enhance critical features while suppressing noise—like computational sunglasses for raw solar data"

2. Long-Short Attention Blocks

   • Local attention for granular magnetic field interactions
   • Global attention for coronal mass ejection tracking

3. Physics-Aware Pretraining

   • Phase 1: 160k training steps for one-step forecasting
   • Phase 2: Autoregressive tuning for multi-hour predictions

The model's 366 million parameters create a generalized solar representation that outperforms task-specific models while enabling parameter-efficient fine-tuning via LoRA adapters.

Benchmark Dominance

Surya delivers significant improvements across critical heliophysics tasks:

Practical Implementation

Developers can immediately leverage Surya's capabilities:

git clone https://github.com/NASA-IMPACT/Surya
cd Surya
uv sync  # Install dependencies
source .venv/bin/activate
python -m pytest tests/test_surya.py  # Validation test

Pretrained models and datasets are available on HuggingFace. Downstream tasks like 24-hour flare forecasting require just single-GPU fine-tuning:

torchrun --nnodes=1 --nproc_per_node=1 --standalone finetune.py

Why This Matters for Earth

Surya isn't just academic—it's critical infrastructure for space weather prediction. Solar flares and coronal mass ejections can:
- Disrupt satellite communications
- Damage power grids
- Endanger astronauts

By unifying multiple prediction tasks into one architecture, Surya enables more accurate, timely warnings. Its zero-shot capabilities mean researchers can immediately probe solar behaviors without collecting new training data—potentially accelerating discoveries about our star's fundamental processes.

As foundation models revolutionize terrestrial AI, Surya proves the same architecture can illuminate our nearest star. This collaboration between NASA and IBM signals a new era where AI doesn't just analyze space data—it helps us comprehend cosmic forces governing our technological existence.

Source: NASA-IMPACT/Surya on GitHub