2026-07-15 –, Thomas Swain Room
Artificial Intelligence (AI) is reshaping meteorological science across two distinct frontiers. On one end, foundation-scale generative models, large-scale distributed training, and massive ensembles push the limits of high-performance computing and big-data orchestration. On the other, a "democratized edge" is emerging, where lightweight, heterogeneous inference workflows broaden access for experimentation. This dual expansion introduces a new class of software challenges spanning distributed training, ensemble-scale orchestration, and efficient, flexible inference pipelines.
This talk will introduce Earth2Studio and PhysicsNeMo from NVIDIA, two software packages designed to enable and scale AI weather forecasting. By exploring their architectures, we will discuss the broader development journey of building AI-driven meteorological tools and share key lessons learned in managing the intersection of high-performance computing, data science and operational reliability.
This is a talk for software engineers, data scientists, and climate researchers navigating the transition from classical simulation to AI-driven meteorology. The following core topics will be presented:
Framework Spotlight: NVIDIA PhysicsNeMo and Earth2Studio
We will introduce and compare two pivotal frameworks from NVIDIA's Earth-2 stack:
- PhysicsNeMo: An open-source Python framework designed for developing AI-physics models at scale. We will discuss its architecture for high-throughput training specifically optimized for weather and climate datasets.
- Earth2Studio: A modular inference and pipeline toolkit. We explore how Earth2Studio allows developers to chain together diverse data sources (ERA5, GFS, satellite) with pre-trained models to create production-ready AI workflows.
Architectural Paradigms in AI Weather
This talk dissects the various model paradigms currently dominating the field and the unique software requirements of each:
- Prognostic Forecast Models: Such as StormScope, FourCastNet or GraphCast, which require stateful time-integration loops that autoregress forward in time, generating forecasts.
- Diagnostic Models: Used for high-resolution downscaling (e.g., CorrDiff) or predicting new products from a forecast system relevant to a particular use case.
- Data Assimilation Models: The bridge between raw satellite/sensor observations and model states, representing an emerging class of AI models accelerating weather and climate data assimilation.
The Challenges of the AI-Weather Stack
Moving from a research notebook to an operational service introduces significant challenges, which this session will address including:
- Data Gravity & Structures: We will discuss the challenges of managing multi-petabyte datasets like ERA5 and the nuances of data formats (Zarr, NetCDF) when moving between high-bandwidth training and low-latency inference.
- Scalability During Training: Designing models must have scalability in mind, navigating both the requirements for data pipelines as well as underlying architectures. State-of-the-art skill and impact often involves high-resolution and/or ensemble training, necessitating advanced parallelism techniques.
- Operational Deployment: Lessons learned in deploying these models into production for users.
- API Standardization & Model Interoperability: We will also discuss the challenges and solutions surrounding offering a large and diverse class of AI models under the same package(s) and providing a unified API for users.
Nicholas Geneva is a Senior Software Engineer in HPC/AI at NVIDIA, specializing in the development of platforms that integrate deep learning with physics and climate science. With over a decade of experience in scientific software development, he currently focuses on developing NVIDIA’s software for Earth-2 to enable AI driven weather / climate prediction for everyone. Nicholas has been a core developer of NVIDIA’s PhysicsNeMo Python packages for the past four years.