-
Notifications
You must be signed in to change notification settings - Fork 27
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* add Ahmed body documentation * fix spelling * update wand instructions * adding vortex shedding doc * AFNO documentation * minor edits * afno getting started * first attempt to document SFNO * graphcast documentation * minor change * address review comments * address more review comments * Update AFNO doc * Update SFNO doc * Update GraphCast doc
- Loading branch information
1 parent
b3dc226
commit 562dc60
Showing
11 changed files
with
572 additions
and
1 deletion.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,119 @@ | ||
# AeroGraphNet for external aerodynamic evaluation | ||
|
||
This example demonstrates how to train the AeroGraphNet model for external aerodynamic | ||
analysis of simplified (Ahmed body-type) car geometries. AeroGraphNet is based on the | ||
MeshGraphNet architecture. It achieves good accuracy on predicting the pressure and | ||
wall shear stresses on the surface mesh of the Ahmed body-type geometries, as well as | ||
the drag coefficient. | ||
|
||
## Problem overview | ||
|
||
To goal is to develop an AI surrogate model that can use simulation data to learn the | ||
external aerodynamic flow over parameterized Ahmed body shape. This serves as a baseline | ||
for more refined models for realistic car geometries. The trained model can be used to | ||
predict the change in drag coefficient,and surface pressure and wall shear stresses due | ||
to changes in the car geometry. This is a stepping stone to applying similar approaches | ||
to other application areas such as aerodynamic analysis of aircraft wings, real car | ||
geometries, etc. | ||
|
||
## Dataset | ||
|
||
Industry-standard Ahmed-body geometries are characterized by six design parameters: | ||
length, width, height, ground clearance, slant angle, and fillet radius. Refer | ||
to the [wiki](https://www.cfd-online.com/Wiki/Ahmed_body) for details on Ahmed | ||
body geometry. In addition to these design parameters, we include the inlet velocity to | ||
address a wide variation in Reynolds number. We identify the design points using the | ||
Latin hypercube sampling scheme for space filling design of experiments and generate | ||
around 500 design points. | ||
|
||
The aerodynamic simulations were performed using the GPU-accelerated OpenFOAM solver | ||
for steady-state analysis, applying the SST K-omega turbulence model. These simulations | ||
consist of 7.2 million mesh points on average, but we use the surface mesh as the input | ||
to training which is roughly around 70k mesh nodes. | ||
|
||
To request access to the full dataset, please reach out to the | ||
[NVIDIA Modulus team]([email protected]). | ||
|
||
## Model overview and architecture | ||
|
||
The AeroGraphNet model is based on the MeshGraphNet architecture which is instrumental | ||
for learning from mesh-based data using GNNs. The inputs to the model are: | ||
|
||
- Ahmed body surface mesh | ||
- Reynolds number | ||
- Geometry parameters (optional, including length, width, height, ground clearance, | ||
slant angle, and fillet radius) | ||
- surface normals (optional) | ||
|
||
Output of the model are: | ||
|
||
- Surface pressure | ||
- Wall shear stresses | ||
- Drag coefficient | ||
|
||
![Comparison between the AeroGraphNet prediction and the | ||
ground truth for surface pressure, wall shear stresses, and the drag coefficient for one | ||
of the samples from the test dataset.](../../../docs/img/ahmed_body_results.png) | ||
|
||
The input to the model is in form of a `.vtp` file and is then converted to | ||
bi-directional DGL graphs in the dataloader. The final results are also written in the | ||
form of `.vtp` files in the inference code. A hidden dimensionality of 256 is used in | ||
the encoder, processor, and decoder. The encoder and decoder consist of two hidden | ||
layers, and the processor includes 15 message passing layers. Batch size per GPU is | ||
set to 1. Summation aggregation is used in the | ||
processor for message aggregation. A learning rate of 0.0001 is used, decaying | ||
exponentially with a rate of 0.99985. Training is performed on 8 NVIDIA A100 | ||
GPUs, leveraging data parallelism. Total training time is 4 hours, and training is | ||
performed for 500 epochs. | ||
|
||
## Getting Started | ||
|
||
The dataset for this example is not publicly available. To get access, please reach out | ||
to the [NVIDIA Modulus team]([email protected]). | ||
|
||
This example requires the `pyvista` and `vtk` libraries. Install with | ||
|
||
```bash | ||
pip install pyvista vtk | ||
``` | ||
|
||
To train the model, run | ||
|
||
```bash | ||
python train.py | ||
``` | ||
|
||
Data parallelism is also supported with multi-GPU runs. To launch a multi-GPU training, | ||
run | ||
|
||
```bash | ||
mpirun -np <num_GPUs> python train.py | ||
``` | ||
|
||
If running in a docker container, you may need to include the `--allow-run-as-root` in | ||
the multi-GPU run command. | ||
|
||
Progress and loss logs can be monitored using Weights & Biases. To activate that, | ||
set `wandb_mode` to `online` in the `constants.py`. This requires to have an active | ||
Weights & Biases account. You also need to provide your API key. There are multiple ways | ||
for providing the API key but you can simply export it as an environment variable | ||
|
||
```bash | ||
export WANDB_API_KEY=<your_api_key> | ||
``` | ||
|
||
The URL to the dashboard will be displayed in the terminal after the run is launched. | ||
Alternatively, the logging utility in `train.py` can be switched to MLFlow. | ||
|
||
Once the model is trained, run | ||
|
||
```bash | ||
python inference.py | ||
``` | ||
|
||
This will save the predictions for the test dataset in `.vtp` format in the `results` | ||
directory. Use Paraview to open and explore the results. | ||
|
||
## References | ||
|
||
- [Learning Mesh-Based Simulation with Graph Networks](https://arxiv.org/abs/2010.03409) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,130 @@ | ||
# MeshGraphNet for transient vortex shedding | ||
|
||
This example is a re-implementation of the DeepMind's vortex shedding example | ||
<https://github.com/deepmind/deepmind-research/tree/master/meshgraphnets> in PyTorch. | ||
It demonstrates how to train a Graph Neural Network (GNN) for evaluation of the | ||
transient vortex shedding on parameterized geometries. | ||
|
||
## Problem overview | ||
|
||
Mesh-based simulations play a central role in modeling complex physical systems across | ||
various scientific and engineering disciplines. They offer robust numerical integration | ||
methods and allow for adaptable resolution to strike a balance between accuracy and | ||
efficiency. Machine learning surrogate models have emerged as powerful tools to reduce | ||
the cost of tasks like design optimization, design space exploration, and what-if | ||
analysis, which involve repetitive high-dimensional scientific simulations. | ||
|
||
However, some existing machine learning surrogate models, such as CNN-type models, | ||
are constrained by structured grids, | ||
making them less suitable for complex geometries or shells. The homogeneous fidelity of | ||
CNNs is a significant limitation for many complex physical systems that require an | ||
adaptive mesh representation to resolve multi-scale physics. | ||
|
||
Graph Neural Networks (GNNs) present a viable approach for surrogate modeling in science | ||
and engineering. They are data-driven and capable of handling complex physics. Being | ||
mesh-based, GNNs can handle geometry irregularities and multi-scale physics, | ||
making them well-suited for a wide range of applications. | ||
|
||
## Dataset | ||
|
||
We rely on DeepMind's vortex shedding dataset for this example. The dataset includes | ||
1000 training, 100 validation, and 100 test samples that are simulated using COMSOL | ||
with irregular triangle 2D meshes, each for 600 time steps with a time step size of | ||
0.01s. These samples vary in the size and the position of the cylinder. Each sample | ||
has a unique mesh due to geometry variations across samples, and the meshes have 1885 | ||
nodes on average. Note that the model can handle different meshes with different number | ||
of nodes and edges as the input. | ||
|
||
## Model overview and architecture | ||
|
||
The model is free-running and auto-regressive. It takes the initial condition as the | ||
input and predicts the solution at the first time step. It then takes the prediction at | ||
the first time step to predict the solution at the next time step. The model continues | ||
to use the prediction at time step $t$ to predict the solution at time step $t+1$, until | ||
the rollout is complete. Note that the model is also able to predict beyond the | ||
simulation time span and extrapolate in time. However, the accuracy of the prediction | ||
might degrade over time and if possible, extrapolation should be avoided unless | ||
the underlying data patterns remain stationary and consistent. | ||
|
||
The model uses the input mesh to construct a bi-directional DGL graph for each sample. | ||
The node features include (6 in total): | ||
|
||
- Velocity components at time step $t$, i.e., $u_t$, $v_t$ | ||
- One-hot encoded node type (interior node, no-slip node, inlet node, outlet node) | ||
|
||
The edge features for each sample are time-independent and include (3 in total): | ||
|
||
- Relative $x$ and $y$ distance between the two end nodes of an edge | ||
- L2 norm of the relative distance vector | ||
|
||
The output of the model is the velocity components at time step t+1, i.e., | ||
$u_{t+1}$, $v_{t+1}$, as well as the pressure $p_{t+1}$. | ||
|
||
![Comparison between the MeshGraphNet prediction and the | ||
ground truth for the horizontal velocity for different test samples. | ||
](../../../docs/img/vortex_shedding.gif) | ||
|
||
A hidden dimensionality of 128 is used in the encoder, | ||
processor, and decoder. The encoder and decoder consist of two hidden layers, and | ||
the processor includes 15 message passing layers. Batch size per GPU is set to 1. | ||
Summation aggregation is used in the | ||
processor for message aggregation. A learning rate of 0.0001 is used, decaying | ||
exponentially with a rate of 0.9999991. Training is performed on 8 NVIDIA A100 | ||
GPUs, leveraging data parallelism for 25 epochs. | ||
|
||
## Getting Started | ||
|
||
This example requires the `tensorflow` library to load the data in the `.tfrecord` | ||
format. Install with | ||
|
||
```bash | ||
pip install tensorflow | ||
``` | ||
|
||
To download the data from DeepMind's repo, run | ||
|
||
```bash | ||
cd raw_dataset | ||
sh download_dataset.sh cylinder_flow | ||
``` | ||
|
||
To train the model, run | ||
|
||
```bash | ||
python train.py | ||
``` | ||
|
||
Data parallelism is also supported with multi-GPU runs. To launch a multi-GPU training, | ||
run | ||
|
||
```bash | ||
mpirun -np <num_GPUs> python train.py | ||
``` | ||
|
||
If running in a docker container, you may need to include the `--allow-run-as-root` in | ||
the multi-GPU run command. | ||
|
||
Progress and loss logs can be monitored using Weights & Biases. To activate that, | ||
set `wandb_mode` to `online` in the `constants.py`. This requires to have an active | ||
Weights & Biases account. You also need to provide your API key. There are multiple ways | ||
for providing the API key but you can simply export it as an environment variable | ||
|
||
```bash | ||
export WANDB_API_KEY=<your_api_key> | ||
``` | ||
|
||
The URL to the dashboard will be displayed in the terminal after the run is launched. | ||
Alternatively, the logging utility in `train.py` can be switched to MLFlow. | ||
|
||
Once the model is trained, run | ||
|
||
```bash | ||
python inference.py | ||
``` | ||
|
||
This will save the predictions for the test dataset in `.gif` format in the `animations` | ||
directory. | ||
|
||
## References | ||
|
||
- [Learning Mesh-Based Simulation with Graph Networks](https://arxiv.org/abs/2010.03409) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,115 @@ | ||
# Adaptive Fourier Neural Operator (AFNO) for weather forecasting | ||
|
||
This repository contains the code used for [FourCastNet: A Global Data-driven | ||
High-resolution Weather Model using Adaptive Fourier Neural | ||
Operators](https://arxiv.org/abs/2202.11214) | ||
|
||
The code was developed by the authors of the preprint: | ||
Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, | ||
Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, | ||
Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, Animashree Anandkumar | ||
|
||
## Problem overview | ||
|
||
FourCastNet, short for Fourier Forecasting Neural Network, is a global data-driven | ||
weather forecasting model that provides accurate short to medium-range global | ||
predictions at 0.25∘ resolution. FourCastNet accurately forecasts high-resolution, | ||
fast-timescale variables such as the surface wind speed, precipitation, and atmospheric | ||
water vapor. It has important implications for planning wind energy resources, | ||
predicting extreme weather events such as tropical cyclones, extra-tropical cyclones, | ||
and atmospheric rivers. FourCastNet matches the forecasting accuracy of the ECMWF | ||
Integrated Forecasting System (IFS), a state-of-the-art Numerical Weather Prediction | ||
(NWP) model, at short lead times for large-scale variables, while outperforming IFS | ||
for variables with complex fine-scale structure, including precipitation. FourCastNet | ||
generates a week-long forecast in less than 2 seconds, orders of magnitude faster than | ||
IFS. The speed of FourCastNet enables the creation of rapid and inexpensive | ||
large-ensemble forecasts with thousands of ensemble-members for improving probabilistic | ||
forecasting. We discuss how data-driven deep learning models such as FourCastNet are a | ||
valuable addition to the meteorology toolkit to aid and augment NWP models. | ||
|
||
FourCastNet is based on the [vision transformer architecture with Adaptive Fourier | ||
Neural Operator (AFNO) attention](https://openreview.net/pdf?id=EXHG-A3jlM) | ||
|
||
![Comparison between the FourCastNet and the ground truth (ERA5) for $u-10$ for | ||
different lead times.](../../../docs/img/FourCastNet.gif) | ||
|
||
## Dataset | ||
|
||
The model is trained on a 20-channel subset of the ERA5 reanalysis data on single levels and | ||
pressure levels that is pre-processed and stored into HDF5 files. | ||
The subset of the ERA5 training data that FCN was trained on is hosted at the | ||
National Energy Research Scientific Computing Center (NERSC). For convenience | ||
[it is available to all via Globus](https://app.globus.org/file-manager?origin_id=945b3c9e-0f8c-11ed-8daf-9f359c660fbd&origin_path=%2F~%2Fdata%2F). | ||
You will need a Globus account and will need to be logged in to your account in order | ||
to access the data. You may also need the [Globus Connect](https://www.globus.org/globus-connect) | ||
to transfer data. The full dataset that this version of FourCastNet was trained on is | ||
approximately 5TB in size. | ||
|
||
## Model overview and architecture | ||
|
||
Please refer to the [reference paper](https://arxiv.org/abs/2202.11214) to learn about | ||
the model architecture. | ||
|
||
## Getting Started | ||
|
||
To train the model, run | ||
|
||
```bash | ||
python train_era5.py | ||
``` | ||
|
||
Progress can be monitored using MLFlow. Open a new terminal and navigate to the training | ||
directory, then run: | ||
|
||
```bash | ||
mlflow ui -p 2458 | ||
``` | ||
|
||
View progress in a browser at <http://127.0.0.1:2458> | ||
|
||
Data parallelism is also supported with multi-GPU runs. To launch a multi-GPU training, | ||
run | ||
|
||
```bash | ||
mpirun -np <num_GPUs> python train_era5.py | ||
``` | ||
|
||
If running inside a docker container, you may need to include the `--allow-run-as-root` | ||
in the multi-GPU run command. | ||
|
||
## References | ||
|
||
If you find this work useful, cite it using: | ||
|
||
```text | ||
@article{pathak2022fourcastnet, | ||
title={Fourcastnet: A global data-driven high-resolution weather model | ||
using adaptive fourier neural operators}, | ||
author={Pathak, Jaideep and Subramanian, Shashank and Harrington, Peter | ||
and Raja, Sanjeev and Chattopadhyay, Ashesh and Mardani, Morteza | ||
and Kurth, Thorsten and Hall, David and Li, Zongyi and Azizzadenesheli, Kamyar | ||
and Hassanzadeh, Pedram and Kashinath, Karthik and Anandkumar, Animashree}, | ||
journal={arXiv preprint arXiv:2202.11214}, | ||
year={2022} | ||
} | ||
``` | ||
|
||
ERA5 data was downloaded from the Copernicus Climate Change Service (C3S) | ||
Climate Data Store. | ||
|
||
```text | ||
Hersbach, H., Bell, B., Berrisford, P., Biavati, G., Horányi, A., Muñoz Sabater, J., | ||
Nicolas, J., Peubey, C., Radu, R., Rozum, I., Schepers, D., Simmons, A., Soci, C., | ||
Dee, D., Thépaut, J-N. (2018): ERA5 hourly data on pressure levels from 1959 to present. | ||
Copernicus Climate Change Service (C3S) Climate Data Store (CDS). 10.24381/cds.bd0915c6 | ||
Hersbach, H., Bell, B., Berrisford, P., Biavati, G., Horányi, A., Muñoz Sabater, J., | ||
Nicolas, J., Peubey, C., Radu, R., Rozum, I., Schepers, D., Simmons, A., Soci, C., | ||
Dee, D., Thépaut, J-N. (2018): ERA5 hourly data on single levels from 1959 to present. | ||
Copernicus Climate Change Service (C3S) Climate Data Store (CDS). 10.24381/cds.adbb2d47 | ||
``` | ||
|
||
Other references: | ||
|
||
[Adaptive Fourier Neural Operators: | ||
Efficient Token Mixers for Transformers](https://openreview.net/pdf?id=EXHG-A3jlM) |
Oops, something went wrong.