Revealing in-plane grain boundary composition features through machine learning from atom probe tomography data