Generative modeling of nucleon-nucleon interactions

A key challenge in ab initio nuclear theory is to develop high-precision models of the nuclear force and propagate the associated uncertainties in many-body calculations of atomic nuclei and nuclear matter. Here we show that modern generative machine learning models can be used to construct novel instances of the nucleon-nucleon interaction when trained on existing potentials from the literature. Specifically, we train the generative model on nucleon-nucleon potentials from chiral effective field theory at three different choices of the resolution scale. We then show that the model can generate novel samples of the nucleon-nucleon potential over a continuous distribution of resolution scales. Finally, we show that the generated potentials are able to produce high-quality nucleon-nucleon scattering phase shifts.