Addressing few-body nuclear systems by Physics-Informed Neural Networks

Recent years have witnessed a growing interest in simulating nuclear systems with deep learning techniques. We propose the use of physics-informed neural networks (PINNs) to address the properties of atomic nuclei subject to realistic nuclear forces. The framework of PINNs constitutes a promising new avenue in deep learning thanks to the possibility of actively exploiting the physical laws and knowledge about the system to shape and train the neural network. Here, we employ PINNs to solve the Schrödinger equation for selected few-body nuclear systems.

We exploit realistic nuclear forces based on chiral effective field theories (χEFT), both in momentum and coordinate spaces, that are consistent with the symmetries of QCD and guarantee the best predictive power in nuclear physics simulations.

We obtain the eigenenergies and corresponding eigenfunctions, with an error for the estimation of the energy below 1 %.

These results pave the way to leverage unsupervised PINNs for simulating the structure and dynamics of more complex nuclei and, in general, to bolster the integration of artificial intelligence techniques for enhancing conventional numerical methods in nuclear physics research.