mindscience/SciAI

• MindSpore SciAI
• • MindSpore SciAI Introduction
  • Model Library
  • Installation Guide
  • • Confirming the System Environment Information
    • Method 1: Install With Pip
    • Method 2: Install From Source Code
    • Installation Verification
  • Quick Start
  • • Method 1: Launch by Advanced API (recommended)
    • Method 2: Launch by Source code
  • Documents
  • Community
  • Contribution
  • Release Notes
  • License

ENGLISH | 简体中文

MindSpore SciAI

• MindSpore SciAI Introduction
• Model Library
• Installation Guide
  • Confirming the System Environment Information
  • Method 1: Install with Pip
  • Method 2: Install From Source Code
  • Installation Verification
• Quick Start
  • Method 1: Launch by Advanced API (recommended)
  • Method 2: Launch by Source code
• Documents
• Community
• Contribution
• Release Notes
• License

MindSpore SciAI Introduction

Based on MindSpore, SciAI is a model library with 60 built-in most frequently used and cited AI4Sci(AI for Science) models, which cover SOTA models from physics-informed (PINNs, DeepRitz, PFNN, etc.) to neural operators (FNO, DeepONet, PDENet), ranking No.1 in the world in terms of coverage.
MindSpore SciAI provides the developers and users with a high-level API (auto configuration, auto instantiation, training and fine-tuning, etc.), allowing an immediate deployment.
With these features, MindSpore SciAI covers a wide variety of scientific computation fields,
including fluid dynamics, electromagnetism, sound, heat, solid and biology, providing developers and users with an efficient and convenient AI4SCI computation platform.

The hierarchy of project SciAI is given below.

├── SciAI
│   ├── cmake               # compilation related
│   ├── docs                # documentation
│   │   ├── FAQ             # frequently asked questions
│   ├── sciai               # SciAI framework
│   │   ├── architecture    # base neural networks modules
│   │   ├── common          # common modules
│   │   ├── context         # context functions
│   │   ├── model           # **AI4SCI high-frequency model library**
│   │   ├── operators       # high order derivatives
│   │   └── utils           # Other auxiliary functions
│   └── tutorial            # tutorial models

Model Library

SciAI model library provides a wide variety of models that are frequently used and cited in scientific computation.
The following table summarizes the current available neural networks and their corresponding domains.

Domain	Network	MindSpore Implementation and Parameters	Ascend	GPU
General Physics	auq_pinns	link	✅	✅
General Physics	cpinns	link	✅	✅
General Physics	deep_hpms	link	✅	✅
General Physics	deep_ritz	link	✅	✅
General Physics	deepbsde	link		✅
General Physics	deeponet	link	✅	✅
General Physics	dgm	link	✅	✅
General Physics	fbsnns	link	✅	✅
General Physics	fpinns	link	✅	✅
General Physics	gradient_pathologies_pinns	link	✅	✅
General Physics	hp_vpinns	link	✅	✅
General Physics	laaf	link	✅	✅
General Physics	mgnet	link	✅	✅
General Physics	multiscale_pinns	link	✅	✅
General Physics	pfnn	link		✅
General Physics	phygeonet	link	✅	✅
General Physics	pi_deeponet	link		✅
General Physics	pinns	link		✅
General Physics	pinns_ntk	link	✅	✅
General Physics	ppinns	link	✅	✅
General Physics	xpinns	link	✅	✅
Hamiltonian Systems	sympnets	link	✅	✅
Fluid Dynamic	hfm	link	✅	✅
Fluid Dynamic	label_free_dnn_surrogate	link	✅	✅
Fluid Dynamic	nsf_nets	link	✅	✅
Fluid Dynamic	*burgers_fno	link	✅	✅
Fluid Dynamic	*burgers_kno	link	✅	✅
Fluid Dynamic	*navier_stokes_fno	link	✅	✅
Fluid Dynamic	*navier_stokes_kno	link	✅	✅
Fluid Dynamic	*navier_stokes_3d_fno	link	✅	✅
Fluid Dynamic	*pde_net	link	✅	✅
Fluid Dynamic	*percnn	link	✅	✅
Elastodynamics	pinn_elastodynamics	link	✅	✅
Thermodynamics	pinn_heattransfer	link	✅	✅
Meteorology	enso	link	✅	✅
Geology	inversion_net	link	✅	✅
Geology	pinn_helmholtz	link	✅	✅
Oceanic Physics	ocean_model	link		✅
Oceanic Physics	pinns_swe	link	✅	✅
Electromagnetism	maxwell_net	link	✅	✅
Electromagnetism	*AD_FDTD_invert_f	link		✅
Electromagnetism	*AD_FDTD_microstrip_filter	link		✅
Electromagnetism	*AD_FDTD_inverse	link		✅
Electromagnetism	*frequency_domain_maxwell	link	✅	✅
Electromagnetism	*frequency_domain_maxwell_3D_dielectric_slab	link	✅	✅
Electromagnetism	*frequency_domain_maxwell_3D_waveguide_cavity	link	✅	✅
Electromagnetism	*meta_auto_decoder	link	✅	✅
Electromagnetism	*pinn_fwi	link	✅	✅
Electromagnetism	*SED_ANN	link	✅	✅
Electromagnetism	*time_domain_maxwell	link	✅	✅
Electromagnetism	*metasurface_holograms	link	✅	✅
Biology	*MEGA-Fold	link (inference) link (training)	✅	✅
Biology	*MEGA-EvoGen	link	✅	✅
Biology	*MEGA-Assessment	link (inference) link (training)	✅	✅
Biology	*ColabDesign	link	✅	✅
Biology	*DeepFRI	link	✅	✅
Biology	*Multimer	link	✅	✅
Biology	*ProteinMPNN	link	✅	✅
Biology	*UFold	link	✅	✅
Biology	*esm-if1	link	✅	✅
Biology	*esm2	link	✅	✅
Biology	*grover	link	✅	✅

Note: the "*" in the model names indicates that these models have already been released at an earlier time by MindSpore
and MindScience.

Installation Guide

Confirming the System Environment Information

Due to the dependency between SciAI and MindSpore, please follow the table below and install the corresponding MindSpore
version from MindSpore Download Guide.

SciAI Version	Branch	MindSpore Version	Python Version
0.1.0	master	>=2.2.0	=3.7

For other dependencies, see requirements.txt,
and they can be installed via the following command.

pip install -r requirements.txt

The supported hardware and operating systems are listed below.

Hardware	OS	Status
Ascend 910	Ubuntu-x86	✅
	Ubuntu-aarch64	✅
	EulerOS-aarch64	✅
	CentOS-x86	✅
	CentOS-aarch64	✅
GPU CUDA 11.1	Ubuntu-x86	✅

Method 1: Install With Pip

This method installs SciAI from .whl package automatically downloaded from MindSpore website,
which does not require the download and compilation of source code.
Note: This installation method is not currently supported.
It will be supported after Mindspore 2.2.0 official release.

pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.2.0/MindScience/sciai/gpu/{arch}/cuda-11.1/sciai-{version}-cp37-cp37m-linux_{arch}.whl -i https://pypi.tuna.tsinghua.edu.cn/simple

• When the network is connected, dependencies of the SciAI installation package are automatically downloaded during
  the .whl package installation. For details about dependencies, see setup.py.
• {version} denotes the version of SciAI. For example, when you are installing SciAI 0.1.0, {version} should
  be 0.1.0.
• {arch} denotes the system architecture. For example, the Linux system you are using is x86 architecture 64-bit,
  {arch} should be x86_64. If the system is ARM architecture 64-bit, then it should be aarch64.

The following table provides the corresponding installation commands to each architecture and Python version.

Device	Architecture	Python	Command
Ascend	x86_64	Python=3.7	pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.2.0/MindScience/sciai/gpu/x86_64/cuda-11.1/sciai-0.1.0-cp37-cp37m-linux_x86_64.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
	aarch64	Python=3.7	pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.2.0/MindScience/sciai/ascend/aarch64/sciai-0.1.0-cp37-cp37m-linux_aarch64.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
GPU	x86_64	Python=3.7	pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.2.0/MindScience/sciai/gpu/x86_64/cuda-11.1/sciai-0.1.0-cp37-cp37m-linux_x86_64.whl -i https://pypi.tuna.tsinghua.edu.cn/simple

Note: If you have other MindScience package(s) installed in your conda or python env, such as MindElec, MindFlow
, MindSponge,
please uninstall the MindScience package(s) in the environment first to avoid pip behavior conflicts.

Method 2: Install From Source Code

1. Clone the source code from the Git repository of MindScience.

   cd ~
   git clone https://gitee.com/mindspore/mindscience.git
   

2. Build SciAI with script build.sh.

   cd mindscience/SciAI
   bash build.sh -j8
   

3. Install the .whl package

   bash install.sh
   

Installation Verification

To verify the installation, run the following commands. If the error message No module named 'sciai' is not displayed,
the installation is successful.

python -c 'import sciai'

Quick Start

The model library SciAI provides two means of model training and evaluation,
providing the developers with flexible and simple usage and development interface.

Method 1: Launch by Advanced API (recommended)

After installation of SciAI:

• Train and Finetune with AutoModel

Users can call the function AutoModel.from_pretrained to get the network models, which are supported in SciAI.
Users can use the function AutoModel.train to train the neural networks, and before training,
users can use AutoModel.update_config to configure the training parameters
or finetune the model by loading the .ckpt file.
The acceptable arguments for AutoModel.update_config depend on the model instantiated.
User can refer to the links Model Library,
column MindSpore Implementation and Parameters for detailed information.

from sciai.model import AutoModel

# obtain the `cpinns` model.
model = AutoModel.from_pretrained("cpinns")
# train the network with default configuration.
# the figures, data and logs generated will be saved in your execution path.
model.train()
# or load a `.ckpt` file.
model.update_config(load_ckpt=True, load_ckpt_path="./checkpoints/your_checkpoint_file.ckpt", epochs=500)
# continue training the model based on the loaded parameters.
model.train()

• Evaluate with AutoModel

Users can evaluate the trained/pre-trained networks with the function AutoModel.evaluate.
This function will load the .ckpt files provided in SciAI by default.
Alternatively, user can load their own .ckpt file with the interface AutoModel.update_config.

from sciai.model import AutoModel

# obtain the `cpinns` model
model = AutoModel.from_pretrained("cpinns")
# load the default ckpt file and evaluate the model
model.evaluate()
# or load the ckpt file provided by the user(Optional)
model.update_config(load_ckpt=True, load_ckpt_path="./checkpoints/your_checkpoint_file.ckpt")
# and then evaluate the model
model.evaluate()

• View training and validation results

Users can find a directory named auto_model_xxx in the script running directory, which contains checkpoints, data
, figures, logs four subdirectories, containing checkpoints, data, pictures, and logs respectively.
Users can view training results in these directories. For example:

├── auto_model_cpinns
│   ├── checkpoints                  # checkpoint files
│   ├── data                         # data files
│   ├── figures                      # plot figures
│   └── logs                         # log files

• 可配置参数:

AutoModel.update_config(**kwargs) 支持修改网络参数，包括学习率、训练周期、混合精度等级、动静态图、数据读取保存路径、checkpoints文件加载路径等，
具体请参考网络模型库中MindSpore实现与网络参数链接。常用的可配置参数见下表：

• Configurable parameters:

AutoModel.update_config(**kwargs) supports modifying network parameters, including learning rate, training epochs,
auto mixed precision level, graph or pynative mode, data loading and saving path, checkpoints file loading path, etc.
For details, please refer to the link in MindSpore implementation and network parameters column
in Model Library. Commonly used configurable parameters are shown in the table below:

parameter	description	default values
save_ckpt	whether save checkpoint or not	true
save_fig	whether save and plot figures or not	true
load_ckpt	whether load checkpoint or not	false
save_ckpt_path	checkpoint saving path	./checkpoints
load_ckpt_path	checkpoint loading path	./checkpoints/model_final.ckpt
load_data_path	path to load data	./data
save_data_path	path to save data	./data
figures_path	figures saving path	./figures
log_path	log saving path	./logs
print_interval	time and loss print interval	10
ckpt_interval	checkpoint saving interval	1000
lr	learning rate	8e-4
epochs	number of epochs	15001
download_data	necessary dataset and/or checkpoints	cpinns
force_download	whether download the dataset or not by force	false
amp_level	MindSpore auto mixed precision level	O3
device_id	device id to set	None
mode	MindSpore Graph mode(0) or Pynative mode(1)	0

• Recover Checkpoints and Dataset

Users may encounter the situation that the automatically downloaded checkpoints or dataset are missing or deleted,
while they can always be re-downloaded by updating the config as follows.

model.update_config(force_download=True)

Method 2: Launch by Source code

Users can clone the whole repository and prepare the environment with the following commands.
Running .env will append the path of SciAI directory to the environment variable PYTHONPATH.

git clone https://gitee.com/mindspore/mindscience
source ./mindscience/SciAI/.env

Then train or evaluate a model in Model Library according to the Quick Start in README.md of the
model. Let's take model cpinns as an example:

• Train the Model:

Users can run script train.py to train the model.

cd ./mindscience/SciAI/sciai/model/cpinns/
python ./train.py [--parameters]

• Finetune the Model with .ckpt file:

python ./train.py --load_ckpt true --load_ckpt_path {your_file}.ckpt [--parameters]

• Evaluate the Model:

Users can run script eval.py to evaluate the trained model. Take cpinns as an example:

python ./eval.py [--parameters]

Configurable parameters [--parameters] include the learning rate, training epochs, data saving and loading paths,
checkpoint loading path and so on.
For details, see the links provided in the MindSpore Implementation and Parameters column
in Model Library.

• Recover Checkpoints and Dataset

Users may encounter with the situation that the automatically downloaded checkpoints or dataset are missing or deleted,
while they can always be re-downloaded with the following solutions:

1. Edit the configuration file config.yaml in the model directory, and set the force_download to true.
2. Append the extra parameter --force_download True when launching train.py or eval.py in the terminal.

Documents

For more details about the installation guides, tutorials, and APIs,
see SciAI Documents.
If you encounter problems with SciAI, please refer to FAQ.

Community

MindSpore Open Governance

Contribution

Make your contribution. For more details, please refer to
out Contributor Wiki

Release Notes

RELEASE

License

Apache License 2.0