Global emulation of atomic nuclei

Computing nuclear observables starting from the fundamental forces between the protons and neutrons in an atomic nucleus is an extremely computationally expensive task, due both to the poor scaling of the quantum many-body problem and the complicated nature of the strong force. This becomes especially prohibitive when one wants to assess the uncertainty of such a calculation coming from its many parameters. In this talk, I will present a novel emulator that allows to quickly predict results of costly calculations. By combining a transformer architecture, bayesian neural networks and a multi-fidelity approach in which coarser approximation can be used to train the model on top of the expensive high-precision calculations, state-of-the art emulation of nuclear properties can be achieve for multiple nuclei simultaneously, with a factor of 10^5 speed up. Moreover, the emulator is able to make predictions for nuclei removed for the training data, allowing one to explore regions of the nuclear chart that are challenging to predict with conventional methods.