Bayesian Modeling & Active Learning

Quantifying the uncertainty of molecular models and the posterior robust prediction of simulation results. Using uncertainty to build informative training datasets step by step in an active learning loop.

Recent Papers:

Active learning graph neural networks for partial charge prediction of metal-organic frameworks via dropout Monte Carlo

S. Thaler, F. Mayer, T. Siby, A. Gagliardi, J. Zavadlav, Npj Comput. Mater. 2024, paper, GitHub

Partial charge prediction for Metal Organic Frameworks. The Dropout Monte Carlo is computationally cheaper than the conventional ensemble approach. At the same time, it provides a reliable estimation of the mean absolute error for unseen MOFs, enabling an efficient active learning strategy.

Scalable Bayesian Uncertainty Quantification for Neural Network Potentials: Promise and Pitfalls

S. Thaler, G. Doehner, J. Zavadlav, J. Chem. Theory Comput. 2023, paper, arXiv

Uncertainty Quantification of out-of-distribution properties with scalable Bayesian methods applicable to Graph Neural Network potentials.

Bayesian selection of coarse-grained models of liquid water

J. Zavadlav, G. Arampatzis, P. Koumoutsakos, Sci. Rep. 2019, paper, arXiv

A data driven evaluation and selection for CG water models through a Hierarchical Bayesian framework.