Orbital-based learning methods are a class of methods using atomic orbital (AO) or molecular orbital (MO) descriptors computed from lower level theory as inputs to learn higher quality molecular properties. This class of methods has emerged as a crucial approach for developing highly accurate electronic structure theories. In our previous work, we have introduced two distinct orbital-based learning approaches (MOB-ML & OrbNet), each tailored for specific scales and utilizing different representations computed from varying computational levels. In this project, we would like to (1) try to unify and improve the two orbital-based learning approaches using Deep Kernel Learning (DKL) as a whole idea to bridge the equivarient graph neural network (EGNN) or graph Transformer as a general representation learner and Gaussian Process Regression (GPR) as a regressor, to fully use date, (2) apply current orbital-based learning as a general potential predictor to real applications in various chemical problems, (3) construct a general pretrained chemical representation via orbital representations using Contrastive learning