Chromatin plays an active role in regulating gene expression, and a detailed structural characterization will be crucial for revealing its structure-function relationship. A high-resolution characterization of chromatin structure and dynamics is challenging to the due to the computational cost associated with its large system size. Using a near-atomistic model that accurately describes protein-DNA interactions, we determined the folding landscape of a key chromatin structural motif, the tetra-nucleosome. A novel deep learning framework was developed to calculate the high-dimensional free energy surface efficiently and to explore the optimal folding coordinate and pathways. The impact of histone modifications and the binding of Polycomb Repressor Complex 2 (PRC2) on chromatin folding was investigated as well.