SurKit is a framework for surrogate models that enables researchers to efficiently construct, train, evaluate, deploy, and execute different kinds of surrogate models across multiple backends, providing an intuitive method to integrate advanced deep learning models into workflows.
- Multi-Model Integration: Integrates a variety of data-driven or physics-constrained neural network surrogate models for solving PDEs (e.g., PINN, FNO, UNet, AFNO, etc.).
- Active Learning: Includes multiple sampling strategies such as Latin Hypercube Sampling (LHS), Sobol sequences, and adaptive sampling.
- Uncertainty Quantification: Provides uncertainty quantification through neural network ensembles and Bayesian neural networks.
- Standardized Interface: Features a unified API design that supports rapid training and allows for customization through the assembly or replacement of cores-level modules.
- Multi-GPU Training: Supports distributed data parallelism for efficient training.
- Multiple Backends: Offers a variety of backends, each with its own unique features and advantages.
SurKit requires some prerequisites including NumPy, Matplotlib, and at least one of the following frameworks:
-
PyTorch: PyTorch>=1.9.0
-
Oneflow: Oneflow>=0.9.0
-
JAX: JAX>=0.3.25, Flax>=0.6.4, Optax>=0.1.4
To Install the stable version of Surkit with pip:
pip3 install surkitTo install the latest development version of Surkit from the source code, you can clone the repository using the following command:
git clone https://github.com/CodexDive/SurKit.gitThe following figure illustrates the performance of Surkit on a CFD-solving task using different neural network architectures and backends. The figure shows comparison between the ground truth solution and the predictions from various models:
Please refer to https://surkit.readthedocs.io/en/latest/index.html.