DeepH-2 is a novel deep learning framework designed to enhance the accuracy and efficiency of density functional theory (DFT) Hamiltonian predictions. The framework integrates local-coordinate transformations and equivariant neural networks, overcoming the limitations of previous models like DeepH and DeepH-E3. DeepH-2 uses an equivariant local-coordinate transformer (ELCT) that reduces the computational complexity from \(O(L^6)\) to \(O(L^3)\), allowing for the inclusion of higher angular momentum features. This reduction in complexity significantly improves the model's performance and efficiency. Through comprehensive experiments on monolayer and bilayer graphene and MoS\(_2\), DeepH-2 demonstrates superior accuracy and efficiency compared to its predecessors, achieving sub-meV prediction accuracy and a 60-fold increase in the number of parameters. The framework's advanced neural network techniques, including equivariant transformers, enable it to handle more complex datasets and pave the way for high-accuracy electronic structure studies of large-scale materials.DeepH-2 is a novel deep learning framework designed to enhance the accuracy and efficiency of density functional theory (DFT) Hamiltonian predictions. The framework integrates local-coordinate transformations and equivariant neural networks, overcoming the limitations of previous models like DeepH and DeepH-E3. DeepH-2 uses an equivariant local-coordinate transformer (ELCT) that reduces the computational complexity from \(O(L^6)\) to \(O(L^3)\), allowing for the inclusion of higher angular momentum features. This reduction in complexity significantly improves the model's performance and efficiency. Through comprehensive experiments on monolayer and bilayer graphene and MoS\(_2\), DeepH-2 demonstrates superior accuracy and efficiency compared to its predecessors, achieving sub-meV prediction accuracy and a 60-fold increase in the number of parameters. The framework's advanced neural network techniques, including equivariant transformers, enable it to handle more complex datasets and pave the way for high-accuracy electronic structure studies of large-scale materials.