Committee neural network potentials

Concept

Instead of a single model, multiple models are trained independently to form a committee that offers:

• Averaging over the predictions improves accuracy
• Disagreement between the committee (measured by the standard deviation of the predictions of the members), provides access to an estimate of the generalisation error
• Substantially reduce overfitting issues
• Active learning strategy known as query by committee (QbC): adding previously unlabeled data with maximal committee disagreement to the training set to systematically improve the model