Physics informed deep learning models for exclusive reactions

With the rapid development of nuclear physics programs aimed at studying the 3D partonic structure of nucleons, the use of artificial intelligence data analysis tools becomes crucial in the extraction of physics information from experimental observables. We present an application of physics-informed machine learning techniques for the extraction of Compton form factors, where a deep neural network learns to map kinematic variables to form factors by fitting against observables measured in deeply virtual Compton scattering experiments. Utilizing state of the art variational autoencoder inverse mappers in combination with uncertainty quantification methods, we demonstrate a solution to the extraction of eight CFFs from a single polarization observable with a quantitative error analysis. This work was funded by the Center for Nuclear Femtography, Southeastern Universities Research Association, Washington D.C. and DOE grant DE-SC0016286.