Conceptual

Partially Bayesian Neural Networks for Active Learning in Materials Science

A partially Bayesian neural network makes only a chosen subset of a network's layers probabilistic while the rest stay deterministic, so calibrated predictive uncertainty can be obtained at a fraction of the cost of sampling every weight. Students learn how the choice of which layers to make probabilistic controls both accuracy and uncertainty calibration, how uncertainty-maximizing acquisition drives active learning over a materials or molecular property space, and how priors initialized from theory-pretrained weights transfer knowledge from simulation to sparse experimental data.