Accelerated Computing for Climate forecast

# Accelerated Computing for Climate forecast: 우리집에 슈퍼컴퓨터가 있다면

이 발표는 2024년 9월 26~28일 제주도에서 열린 2024 Climate Tech Summit 에서, AI 를 사용한 국지적/전반적 기후예측 모델에 대한 발전 과정과 현재 모델들에 대한 내용을 다룬 발표입니다. 시뮬레이션 기반의 예측 모델의 엄청난 계산량으로 인한 비용 부담을 AI 앙상블을 이용해서 해결하는 아이디어와, 그 응용 사례 중 하나인 기후 모델이 2022년부터 발전해 온 과정을 소개합니다. 구글, 마이크로소프트, 엔비디아, 화웨이등의 모델과, 대한민국에서 진행된 AI 기반 기상 예측 연구에 대해서도 소개합니다.

This presentation was delivered at the 2024 Climate Tech Summit held in Jeju Island from September 26-28, 2024. It covers the development process and current models of AI-based local and global climate prediction models. The presentation introduces the idea of using AI ensembles to address the cost burden caused by the enormous computational requirements of simulation-based prediction models, and showcases the evolution of climate models—one of the key applications—since 2022. It also introduces models from Google, Microsoft, NVIDIA, and Huawei, as well as AI-based weather prediction research conducted in South Korea.

Transcript

3. 1980 1990 2000 2010 2020 CPU Stock data Online shopping Financial SNS data 216 215 214 217 218
4. Training w/ Synthetic Data Simulator-based Digital Twin Generating Synthetic Data AI-generated synthetic data Real-time or static Extreme edge conditions Simulating without simulation Prediction by inference Result prediction model Training with empirical / synthetic data
5. Tesla AI Day 2021. https://www.youtube.com/watch?v=j0z4FweCy4M&t=6116s
6. Tesla AI Day 2021. https://www.youtube.com/watch?v=j0z4FweCy4M&t=6116s
7. • NVIDIA DriveSim (2021) • NVIDIA self-driving digital twin • NVIDIA Orin + Hyperion 8 Platform
8. Biomedical Synthesis Data with Generative Models [1] Chen, R.J., Lu, M.Y., Chen, T.Y. et al. Synthetic data in machine learning for medicine and healthcare. Nat Biomed Eng 5, 493–497 (2021). https://doi.org/10.1038/s41551-021-00751-8 [2] D'Amico et al., Synthetic data generation by artificial intelligence to accelerate research and precision mecine in hematology, JCO Clin Cancer Inform Jun:7:e2300021 (2023) doi: 10.1200/CCI.23.00021.
9. • @supercomputing23 • High-Performance Computing • Traditionally, GPU-based computation hasn't gotten the attention it deserves • Numerical analysis: precision is key • GPU: float16, bf16, INT8… • Numerical simulations: FP64, (sometimes) FP128… • Accuracy issues caused by error rate accumulation Bypassing simulations ”GraphCast: AI model for faster and more accurate global weather forecasting” (2023) 25
10. • The beginning of change • Numerical analysis is not the end goal of HPC • A New Approach in Applications: Getting Simulation Results Without Simulation • AI-based result inference • AlphaFold, GNoME • Replacing Simulation with Predictive Models • Adoption • Meteorology, hydrodynamics, dynamics, materials science, and more... Bypassing simulations 26
11. • Why Weather Forecasting is Crucial for Climate Change Preparedness • Anticipates catastropic weather events (floods, droughts, storms) • Enables proactive disaster management and response • Supports adaptive agriculture and food security • Aids in urban planning and infrastructure development • Facilitates effective resource management (water, energy) • Informs long-term climate adaptation strategies • Enhances public awareness and preparedness Importance of Weather Forecasting
12. Paradigm shift • FourCastNet (Feb. 2022) • First impactful demonstration of the potential of AI-powered weather prediction • High resolution / Fast prediction speed (2 sec. for 1 week forecast) • Enables ensemble-based prediction • Pangu-Weather (Huawei, Jul. 2023) • Up to 10000x speedup • Training • 37 year-ERA5 data • 192 NVIDIA V100 cards Weather/Climate Prediction by AI https://arxiv.org/abs/2202.11214 https://www.nature.com/articles/d41586-023-02084-9 https://research.nvidia.com/publication/2024-08_kilometer-scale-convection-allowing-model-emulation-using-generative-diffusion Beyond simulation • NeuralGCM (Google, May 2024) • GraphCast (Google, 2023) • Aurora (Microsoft, Apr. 2024) • Iteration of ClimaX (Jan. 2023) • Up to 500,000% speed-up / 10 day prediction • StormCast (NVIDIA, Sep. 2024) • Earth-2 (NVIDIA, 2022) and CorrDiff (NVIDIA, June 2024) • Cost reduction: USD 3M -> USD 60k (2%) • Resolution: 25km -> 2km
13. • AlphaWeather (NIMS & KAIST, 2021~) • RADAR image based • KARINA (KIST, July 2024) • Training • 37 year-ERA5 fine-grained data • 4 NVIDIA A100 GPUs (12 hours) • Saving training resource • Divide into 10x larger grids for efficiency Weather/Climate Prediction by AI in Korea
14. • AI rapidly replacing traditional high-performance computing simulations, reducing costs and time • AI inference costs significantly lower than traditional simulation methods • Ensemble calculations enhancing AI prediction accuracy, addressing potential discrepan cies • Weather forecasting: a prime example of high-performance computing predictions • Since 2022, AI-based forecasting dramatically reducing prediction costs in meteorology • This shift enables more accessible and localized weather predictions, facilitating better a daptation to changing climate conditions Conclusion: AI Revolution in Weather/Climate Forecasting