QUiNN is a Python library that is centered around wrapper classes to PyTorch NN modules enhancing them with predictive uncertainty capabilities. It currently includes MCMC (adaptive and Hamiltonian), Variational Inference (Bayes-by-backprop), Ensemble-based (Deep Ensembles, Randomized MAP Sampling) and Laplace approximation methods. Besides, QUiNN includes residual neural networks (ResNets) with depth-parameterized weights shown to work better for regression problems.

QUiNN has been employed to develop uncertainty-augmented ResNet surrogates for outputs of the DOE E3SM Land model. It is also currently assisting minima-preserving NN construction in DOE BES Exascale Catalytic Chemistry (ECC) project.