SEAVEA 0.24.2

• Notebook tutorial: create small closure model for 2D Navier Stokes (part of NWO ENW-M1 project ``Learning small closure models for large multiscale problems"): https://github.com/wedeling/EasySurrogate/blob/master/tutorials/reduced_Navier_Stokes/
• Generative Adversarial Networks: generative modelling with ANNs, notebook tutorial: https://github.com/wedeling/EasySurrogate/tree/master/tutorials/GAN
• Batch normalization: to avoid covariate shift during training, notebook tutorial: https://github.com/wedeling/EasySurrogate/tree/master/tutorials/batch_normalization

SEAVEA v0.19

Some minor changes have been introduced in the ANN / Deep active subspace methods:

• made it possible to select varying numbers of neurons, to create "constrained" hidden layers. This is used to make a deep-active subspace type of surrogate without orthonormality embedded in network architecture. Instead, orthonormality is computed in a post-processing procedure. Explained in updated tutorials.

Updated tutorials:

• 2 epidemiological deep-active subspace Jupyter notebook tutorials (HIV and COVID 19): https://github.com/wedeling/deep_active_subspace_data

M42 release

New features:

• Separate library for computing the exact derivatives of Gram-Schmidt vectors: https://github.com/wedeling/Gram_Schmidt_Derivatives

Updates:

• Added non-linear activation possibility to deep-active subspace layer

Tutorials:

• 2 epidemiological deep-active subspace Jupyter notebook tutorials (HIV and COVID 19): https://github.com/wedeling/deep_active_subspace_data
• Jupyter notebook tutorial on use and validity the Gram-Schmidt derivatives: https://github.com/wedeling/Gram_Schmidt_Derivatives/blob/main/check_derivatives.ipynb

M36 release

This release is part of the M36 release of the VECMA toolkit. New features:

• Gaussian Process surrogate + tutorial
• Deep active subspace surrogate + tutorial

• Coupling between EasyVVUQ and EasySurrogate, allowing the former to generate training data for the latter + tutorial

0.15 (Beta)

New surrogate methods are added:

• Kernel mixture networks: neural-network based conditional probability density functions
• Vanilla artificial neural networks for comparison purposes.

Tutorials on these networks are also added. A tutorial on coupling surrogates with MUSCLE3 is present as well.

EasySurrogate 0.1 (beta)

First Beta release of EasySurrogate

• Mimics the 'Campaign' structure of EasyVVUQ
• Quantized Softmax Networks are available, which are stochastic, neural-network based surrogates used for conditional resampling of reference data
• Reduced surrogates are available, which are used to compress the size of training data by orders of magnitude, while retaining accuracy for spatially-integrated quantities of interest.
• Contains a general tutorial, and one for each type of available surrogate method.