.svg){kind=spacer}
Details
- NVIDIA and partners released ClimSim-Online, a framework for developing hybrid physics-machine learning climate models.
- The consortium includes NVIDIA Earth 2, international climate modelers, and a Columbia University-based NSF center.
- It uses the ClimSim dataset to train ML models that replace expensive cloud-resolving simulations, accelerating climate modeling by orders of magnitude.
- Current climate models cannot fully capture small-scale processes; CRMs are accurate but too computationally intensive for long simulations.
- The containerized workflow allows plug-and-play ML integration into the E3SM climate simulator, and recent breakthroughs achieved multi-year stable simulations with physics-informed constraints.
Impact
ClimSim-Online democratizes high-resolution climate simulation, potentially reducing uncertainties in projections critical for policy. By merging AI with physics, it overcomes computational barriers, enabling faster and more accurate long-term forecasts. This could accelerate climate science and inform global adaptation strategies.
