Quantifying uncertainty of nuclear properties within machine learning frameworks

In this work, we use deep learning technique to learn properties of atomic nuclei while accounting for the underlying neural network uncertainty that arises from the complex parameter manifold. While ensemble uncertainties offer a first insight into the confidence of machine learning predictions, they are computationally intractable in all but the simplest problems. We instead apply an approximation of the ensemble uncertainty that can be computed both in simple problems as well as in complex machine learning architectures. In this demonstrative calculation, we use nuclear masses computed by density functional theory (DFT). We train on a subset of the computed mass table and examine the quality of the predicted masses, both in interpolative and extrapolative regimes, as well as the resulting ensemble uncertainty within this approximation scheme.