5. Using AI for simulations – How?
AI for simulations
• AI has already disrupted the way we think of computation in other domains and mapping to AI unleashes parallelism
• Doing once vs repetitive – learn once and infer over and over
*HPC+AI ってよく聞くけど結局なんなの, 山崎,
GTC2022 テクニカルフォローアップセミナー
6. NVIDIA Modulus
Platform for developing Physics ML surrogate model
• A training and inference pipeline - using
Physics (governing equations) and Data
(simulation/observations)
• Customizable and scalable platform
• Higher level of abstraction for domain
experts
• Build generalized AI surrogates for
parameterized domain
• Near real-time high-fidelity simulation
Modulus EULA – It’s free.
7. Developing digital twins for weather, climate, and energy [S41823]
12
Physics-ML categorization
Physics
Data
Fully data
driven
Inductive
bias
Physics
constrained
Fully physics
driven
12. NVIDIA'S Earth -2: Digital Twins For Weather and Climate [A41326]
Earth-2 Began by Exploring
Data-Driven Weather Prediction
Scope Global, Medium Range
Model Type Full-Model AI Surrogate
Architecture AFNO (Adaptive Fourier Neural Op.)
Resolution: 25km
Training Data: ERA5 Reanalysis
Initial Condition GFS / UFS
Inference Time 0.25 sec (2-week forecast)
Speedup vs NWP O(104-105)
Power Savings O(104)
FourCastNet
13. NVIDIA'S Earth -2: Digital Twins For Weather and Climate [A41326]
Deuben & Bauer (2018), 6 , 60x30, 1.8K pixels, MLP
WeatherBench, Rasp et al. (2020). 5.625 , 64x32, 2K pixels, CNN
Weyn et al. (2019), 2.5 N.H only, 72x36, 2.6k pixels, ConvLSTM
DLWP, Weyn et al. (2020). 2 , 16K pixels, Deep CNN on Cubesphere/(2021) ResNet
FourCastNet, Pathak et al. (2022), 0.25 , ~1,000,000 Pixels, ViT+AFNO
GNN, Keisler et al. (2022), 1 , 64,000 Pixels, Graph Neural Networks
FourCastNet: A new data-driven weather predictor of unprecedented resolution
14. NVIDIA'S Earth -2: Digital Twins For Weather and Climate [A41326]
We train FCN on ambitious amounts of data on large machines
Thanks to full-stack AI + HPC expertise we train on a growing amount of the world's petabytes of past weather data.
Time to solution decreased from 24+ hours to 67 minutes with model and data parallelism
FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators, Kurth et al. (2022), https://arxiv.org/abs/2208.05419
15. NVIDIA'S Earth -2: Digital Twins For Weather and Climate [A41326]
FCN skill improving with training ambition.
Could it one day outperform deterministic models? We don't yet know the limit.
Skill gap reduced by more than half
w.r.t IFS gold standard
Skill gap reduced by more than half
w.r.t IFS gold standard
Acronym Alert:
ACC: Anomaly Correlation Coefficient (metric of weather skill)
IFS: The Integrated Forecast System, a gold standard weather model
FCN: FourCastNet, our digital twin of weather.
16. Open-Source FourCastNet
Join us in pushing the frontiers of data-driven numerical weather prediction
https://github.com/NVlabs/FourCastNet
17. FourCastNet implementation using Modulus
Included in Modulus examples
https://docs.nvidia.com/deeplearning/modulus/user_guide/neural_operators/fourcastnet.html
18. WIND TURBINE WAKE OPTIMIZATION —
SIEMENS GAMESA
Use Case
§ Developing optimal engineering wake models to optimize wind
farm layouts
§ Simulating the effect that a turbine might have on another when
placed in close proximity
Challenges
§ Generating high-fidelity simulation data from Reynolds-averaged
Navier-Stokes (RANS) or Large Eddy Simulations (LES) can take
over a month to run, even on a 100-CPU cluster.
Solution
§ NVIDIA Omniverse and Modulus enable accurate, high-fidelity
simulations of the wake of the turbines, using low-resolution
simulations as inputs and applying super resolution using AI.
NVIDIA Solution Stack
§ Hardware: NVIDIA A100, A40, RTX 8000 GPUs
§ Software: NVIDIA Omniverse, NVIDIA Modulus
Outcome
§ Approximately 4,000X speedup for high-fidelity simulation
§ Optimizing wind farm layouts in real-time increases overall
production while reducing loads and operating costs.
Demo
19. WIND TURBINE WAKE OPTIMIZATION — SIEMENS GAMESA
Super resolution using AI
https://www.youtube.com/watch?v=mQuvYQmdbtw
20. Super Resolution Net and pix2pixHD Net are now available in Modulus
Included in Modulus examples
https://docs.nvidia.com/deeplearning/modulus/user_guide/intermediate/turbulence_super_resolution.html
23. Learn how to use NVIDIA Modulus
https://github.com/openhackathons-org/gpubootcamp/blob/master/hpc_ai/PINN/English/python/Start_Here.ipynb
24. User Guide (ex. Defining custom PDEs)
NVIDIA Modulus User Guide
https://docs.nvidia.com/deeplearning/modulus/user_guide/foundational/1d_wave_equation.html#writing-custom-pdes-and-boundary-initial-conditions
25. Ask questions at NVIDIA Modulus technical forum
https://forums.developer.nvidia.com/c/physics-simulation/modulus-physics-ml-model-framework/technical-support/453
26. Contribute to NVIDIA Modulus on GitLab!
*Need to submit GitLab repository access request from https://developer.nvidia.com/modulus-downloads
27. Summary
• NVIDIA Modulus is platform for developing Physics-ML surrogate model
• Physics-ML use cases
• FourCastNet
https://arxiv.org/abs/2208.05419, https://github.com/NVlabs/FourCastNet
• Super Resolution
https://blogs.nvidia.com/blog/2022/03/22/siemens-gamesa-wind-farms-digital-twins/
• How to get started
• Download
https://developer.nvidia.com/modulus
• Learn
https://docs.nvidia.com/deeplearning/modulus/index.html
https://github.com/openhackathons-org/gpubootcamp/blob/master/hpc_ai/PINN/English/python/Start_Here.ipynb
• Ask questions
https://forums.developer.nvidia.com/c/physics-simulation/modulus-physics-ml-model-framework/technical-
support/453
• Get involved
https://developer.nvidia.com/modulus-downloads