Scientific Machine Learning With NVIDIA PhysicsNeMo
summary
Shows a workflow that converts OpenFOAM simulations into a trained neural operator that predicts surface pressure, temperature, and velocity in seconds. Uses OpenFOAM with Latin Hypercube Sampling to generate training data, trains NVIDIA PhysicsNeMo DoMINO (9.6M parameters) which encodes 3D geometry as a signed distance field, and deploys models in a web UI called FlowForge end-to-end platform; Ahmed body results report pressure error under 0.8% and wall shear stress error ≈0.0052.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:25What are simulations
- 2:49Simulations in F1 history
- 4:55Industries using simulations
- 5:57AI surrogate modeling need
- 7:22Physics-informed neural networks
- 9:09Neural operators for simulations
- 9:52PhysicsNeMo framework
- 11:19FlowForge platform demo
- 11:59Data ingestion in FlowForge
- 13:02Analyzing data distribution
- 13:45Visualizing 3D geometries
- 14:27Training models easily
- 15:30Configuring hyperparameters
- 15:51Monitoring training runs
- 16:33Running inference
- 17:17Visualizing results with ParaView
- 18:01Comparing model performance
- 18:43Results on Ahmed body CFD
- 20:29Key takeaways
- 21:47Q&A
related talks
01h 11mHidden Physics Models: Machine Learning of Nonlinear Partial Differential Equations
29mBuild Customize and Deploy LLMs At-Scale on Azure with NVIDIA NeMo
33mReproducible Molecular Simulations with the Help of Python
25mConvergence of AI, Simulations and HPC
34mData-Driven Solutions in Engineering
36mKeynote: Scientific Discovery: From the Lab Bench to the GPU
11mDeploying Physics Models at the Edge: Quantized AI for Weather & Fluid Simulation
01h 04mNonlinear Model Reduction: Using ML to Enable Rapid Simulation of Extreme-Scale Physics Models