Software and Research Platforms
Figure: Sea surface temperature field from CESM-HR as submitted to HighResMIP. The right panels show a blow-up of the Western North Atlantic region. Units are in degrees Celsius.
NSF NCAR supports a shared, AI-optimized software ecosystem that enables reproducible, scalable Earth system science workflows. Our software environments are designed to integrate seamlessly with NSF NCAR’s HPC systems and AI-ready data pipelines, helping researchers develop, run, and reproduce reliable AI-enabled research.
Powerful hardware alone is not enough. To accelerate scientific discovery, NSF NCAR provides integrated frameworks, tools, and user environments that support the full AI/ML lifecycle — from training and validation to inference and deployment.
AI Frameworks
NSF NCAR maintains optimized AI frameworks on its GPU-accelerated HPC systems to support distributed, large-scale model development.
Most AI research leverages widely adopted frameworks such as PyTorch, TensorFlow, and JAX (including Haiku and Flax). These packages are carefully built and tuned for compatibility with GPU libraries, high-speed interconnects, and multi-node distributed training environments, ensuring researchers can scale efficiently without managing complex software dependencies.
Beyond general-purpose deep learning libraries, NSF NCAR supports domain-specific AI tools designed for the complexity of Earth system science.
NSF NCAR’s CREDIT package provides end-to-end workflows for training and deploying large AI-based Earth system emulators. CREDIT integrates directly with GPU-accelerated HPC resources and AI-ready data pipelines such as GDEX, enabling streamlined access to analysis-ready datasets and scalable training workflows.
Earth2Studio and Emerging Tools
The broader ecosystem includes tools such as NVIDIA Earth2Studio, an open-source framework supporting emulator inference workflows for Earth system applications. NSF NCAR continues to evaluate and integrate emerging AI tools that advance community science needs.
User-Friendly Platforms
To reduce barriers to entry, NSF NCAR provides interactive environments such as NSF NCAR JupyterHub, offering push-button access to supercomputing resources.
Pre-configured user environments include widely used Earth system science software packages and support both internally developed and externally developed AI models. These platforms allow researchers to prototype, analyze, and deploy models without needing to manage infrastructure or software configuration.
Access to these software environments requires a computing allocation: