This is a first beta release of EasySurrogate, a toolkit designed to facilitate the creation of surrogate models for multiscale simulations. The development of this software is funded by the EU Horizon 2020 Verified Exascale Computing for Multiscale Applications (VECMA) project.
After cloning the repository, EasySurrogate, along with the requirements, can be installed via:
python3 setup.py install --user
An overaching `Campaign' structure of creating surrogates, similar to EasyVVUQ, the VECMA toolkit component for forward uncertainty propagation.
Quantized Softmax Networks: a neural network surrogate based on conditional resampling of reference data. The corresponding article can be found here.
Kernel Mixture Networks: nonparametric estimation of conditional probability densities using neural networks. Unlike the Quantized Softmax Surrogate, no resampling of reference data is performed, and the probability density is continuous rather than discrete.
Reduced surrogates: a data compression technique used to reduce the size of the training data down by several order of magnitude, while retaining accuracy for spatially integrated quantities of interest. The corresponding article can be found here. This has only been tested on problems with 2 spatial dimensions. Will be generalized in a subsequent release.
Standard artificial neural networks, used for regression with (time-lagged) features.
It is possible to couple surrogates to the macroscopic model via MUSCLE3, the third incarnation of the Multiscale Coupling Library and Environment, see the corresponding tutorial.
The following tutorials can be found in the
/General: a general introduction to EasySurrogate.
/L96_QSN: Quantized Softmax Network (QSN) surrogates for atmospheric model equations. In this tutorial, the subgrid-scale term of the Lorenz96 equations is replaced by a QSN surrogate.
/L96_KMN: Kernel Mixture Network (KMN) surrogates for atmospheric model equations. In this tutorial, the subgrid-scale term of the Lorenz96 equations is replaced by a KMN surrogate.
/L96_ANN: Artificial Neural Network (ANN) surrogates for atmospheric model equations. In this tutorial, the subgrid-scale term of the Lorenz96 equations is replaced by an ANN surrogate.
/gray_scott: reduced surrogates for a reaction diffusion equation. This applies the data compression technique to the two-dimensional gray-scott equations.
/gray_scott_muscle: this is basically the same tutorial as above, only the coupling between the reduced surrogates and the reaction diffusion equations is performed using MUSCLE3.