Skip to content

Kontinua: The Open-Core Physics AI Platform

Welcome to Kontinua, the definitive open-source SDK and managed cloud platform for training, benchmarking, and deploying foundation models for physics.

Kontinua provides seamless access to 15TB of curated, analysis-ready simulation data across 16 physical domains—including biological systems, fluid dynamics, acoustic scattering, and magneto-hydrodynamics. Whether you are building surrogate models locally or orchestrating massive distributed training jobs on Kontinua Cloud, our platform provides the tools you need.

Academic Origins: Kontinua's 15TB dataset is powered by The Well, originally developed by the Polymathic AI organization. We are proud to build upon their open-source research and dual-license our SDK under the Apache 2.0 license.

Using the Kontinua SDK

Once the Kontinua package is installed, you can effortlessly load our curated datasets into your PyTorch training pipelines.

from kontinua.data import WellDataset
from torch.utils.data import DataLoader

trainset = WellDataset(
    well_base_path="path/to/base",
    well_dataset_name="name_of_the_dataset",
    well_split_name="train"
)
train_loader = DataLoader(trainset)

for batch in train_loader:
    ...

For more information regarding the interface, please refer to the API and the tutorials.

Installation

We recommend creating a new Python (>=3.10) environment to install Kontinua.

python -m venv path/to/env
source path/to/env/activate/bin

From PyPI

The Kontinua package can be installed directly from PyPI.

pip install kontinua

From Source

It can also be installed from source. For this, clone the repository and install the package with its dependencies.

git clone https://github.com/KontinuaAI/kontinua
cd kontinua
pip install .

Depending on your acceleration hardware, you can specify --extra-index-url to install the relevant PyTorch version. For example, use:

pip install . --extra-index-url https://download.pytorch.org/whl/cu121

Surrogate Modeling Dependencies

If you want to run the model training benchmarks, install the additional dependencies:

pip install kontinua[benchmark]

Downloading the Data

The datasets range between 6.9GB and 5.1TB of data each, for a total of 15TB for the full collection. Ensure that your system has enough free disk space.

Once kontinua is installed, you can use the CLI to download any domain dataset.

kontinua-download --base-path path/to/base --dataset active_matter --split train

Streaming from Hugging Face

Most of the datasets are also hosted on Hugging Face. Data can be streamed directly from the hub using the following code.

from kontinua.data import WellDataset
from torch.utils.data import DataLoader

# The following line may take a couple of minutes to instantiate the datamodule
trainset = WellDataset(
    well_base_path="hf://datasets/KontinuaAI/",  # access from HF hub
    well_dataset_name="active_matter",
    well_split_name="train",
)
train_loader = DataLoader(trainset)

Surrogate Modeling & Training

The repository provides tools for training foundation models and surrogate architectures on the different physical domains. State-of-the-art models are implemented in models, while dataset classes handle the raw data.

The training framework uses hydra to instantiate various classes (e.g., dataset, model, optimizer) from configuration files.

For instance, to run the training script of the default FNO architecture on the active matter dataset:

cd kontinua/benchmark
python train.py experiment=fno server=local data=active_matter

⚡ Kontinua Cloud (Managed Platform)

While the SDK provides everything you need to train models locally, Kontinua Cloud takes physics AI to production. Our managed platform offers: - Serverless Fine-tuning: Train foundation models on our distributed GPU clusters. - REST API Inference: Deploy trained surrogate models and run inference via secure REST API endpoints. - Data Integration: Connect your proprietary simulation data seamlessly.

Sign up for the Kontinua Cloud Waitlist or read our Cloud Documentation.

Academic Citation

If you find the underlying 15TB dataset useful for your academic research, please consider citing the original paper by Polymathic AI:

@article{ohana2024well,
  title={The well: a large-scale collection of diverse physics simulations for machine learning},
  author={Ohana, Ruben and McCabe, Michael and Meyer, Lucas and Morel, Rudy and Agocs, Fruzsina and Beneitez, Miguel and Berger, Marsha and Burkhart, Blakesly and Dalziel, Stuart and Fielding, Drummond and others},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={44989--45037},
  year={2024}
}