NVIDIA PhysicsNeMo Examples

Introduction

This repository provides sample applications demonstrating use of specific Physics-ML model architectures that are easy to train and deploy. These examples aim to show how such models can help solve real world problems.

Introductory examples for learning key ideas

Use case	Concepts covered
Darcy Flow	Introductory example for learning basics of data-driven models on Physics-ML datasets
Darcy Flow (Data + Physics)	Data-driven training with physics-based constraints
Lid Driven Cavity Flow	Purely physics-driven (no external simulation/experimental data) training
Vortex Shedding	Introductory example for learning the basics of MeshGraphNets in PhysicsNeMo
Medium-range global weather forecast using FCN-AFNO	Introductory example on training data-driven models for global weather forecasting (auto-regressive model)
Lagrangian Fluid Flow	Introductory example for data-driven training on Lagrangian meshes
Stokes Flow (Physics Informed Fine-Tuning)	Data-driven training followed by physics-based fine-tuning

Domain-specific examples

The several examples inside PhysicsNeMo can be classified based on their domains as below:

NOTE: The below classification is not exhaustive by any means! One can classify single example into multiple domains and we encourage the users to review the entire list.

NOTE: * Indicates externally contributed examples.

CFD

Use case	Model	Transient
Drag prediction - External Aero	MeshGraphNet, UNet, DoMINO, FigConvNet, Transolver	NO
Drag prediction - External Aero - Mixture of Experts	MoE Model	NO
Navier-Stokes Flow	RNN	YES
Gray-Scott System	RNN	YES
Lagrangian Fluid Flow	MeshGraphNet	YES
Darcy Flow (Data + Physics Driven) using DeepONet approach	FNO (branch) and MLP (trunk)	NO
Darcy Flow (Data + Physics Driven) using PINO approach (Numerical gradients)	FNO	NO
Magnetohydrodynamics using PINO (Data + Physics Driven)*	FNO	YES
Shallow Water Equations using PINO (Data + Physics Driven)*	FNO	YES
Shallow Water Equations using Distributed GNNs	GraphCast	YES
Vortex Shedding with Temporal Attention	MeshGraphNet	YES
Data Center Airflow	3D UNet	NO
Fluid Super-resolution*	Denoising Diffusion Probablistic Model	YES
Pre-trained DPOT for Navier-Stokes*	Denoising Operator Transformer	YES
Fine-tuning of DoMINO NIM	DoMINO	NO
Transolver for External Aerodynamics on Irregular Meshes	Transolver	NO

Weather

Use case	Model
Medium-range global weather forecast using FCN-SFNO	FCN-SFNO
Medium-range global weather forecast using GraphCast	GraphCast
Medium-range and S2S global weather forecast using DLWP	DLWP
Coupled Ocean-Atmosphere Medium-range and S2S global weather forecast using DLWP-HEALPix	DLWP-HEALPix
Diagonistic (Precipitation) model using AFNO	AFNO
Unified Recipe for training several Global Weather Forecasting models	AFNO, FCN-SFNO, GraphCast
Generative Correction Diffusion Model for Km-scale Atmospheric Downscaling	CorrDiff
StormCast: Generative Diffusion Model for Km-scale, Convection allowing Model Emulation	StormCast
Medium-range global weather forecast using Mixture of Experts	MoE Model
Generative Data Assimilation of Sparse Weather Observations	Denoising Diffusion Model
Flood Forecasting	GNN + KAN
Temporal Interpolation of Weather Forecasts	ModAFNO

Structural Mechanics

Use case	Model
Deforming Plate	MeshGraphNet
Machine Learning Surrogates for Automotive Crash Dynamics	Transolver, MeshGraphNet

Healthcare

Use case	Model
Cardiovascular Simulations*	MeshGraphNet
Brain Anomaly Detection	FNO

Additive Manufacturing

Use case	Model
Metal Sintering Simulation*	MeshGraphNet

Molecular Dymanics

Use case	Model
Force Prediciton for Lennard Jones system	MeshGraphNet

Geophysics

Use case	Model
Diffusion model for full-waveform inversion	UNet, Global Filter Net
Reservoir Simulation using X-MeshGraphNet	MeshGraphNet

Generative

Use case	Model
TopoDiff*	Conditional diffusion-model

Active Learning

1. Classify the famous two-moons data distribution using Active learning
2. Active Learning for Surface-CFD Aerodynamic Surrogates

Additional examples

Physics-informed training examples (PINNs, PINO, physics-informed fine-tuning) use the physicsnemo.sym module. Install with pip install "nvidia-physicsnemo[sym]".

NVIDIA support

In each of the example READMEs, we indicate the level of support that will be provided. Some examples are under active development/improvement and might involve rapid changes. For stable examples, please refer the tagged versions.

Feedback / Contributions

We're posting these examples on GitHub to better support the community, facilitate feedback, as well as collect and implement contributions using GitHub issues and pull requests. We welcome all contributions!