Deep learning for strongly correlated systems in quantum chemistry
In this project we plan to develop the algorithm mapping a structure to properties at domain-based local pair natural orbitals (DLPNO) multireference coupled cluster (DLPNO-MRCC) level and DLPNO tailored coupled cluster (DLPNO-TCC) level corrected by density matrix renormalization group (DMRG). Simultaneously, we will develop a machine learning algorithm predicting the most entangled pairs of orbitals, based on precomputed set of one-electron, mono and bicentric integrals. It could be seen as an automatic selection of the active space. We believe that the algorithm trained on small and medium-size systems could exploit learned information also for larger systems, for which is the selection of active space for CAS methods (like DMRG) difficult. Without any significant effort we will also develop the algorithm mapping integrals to a property at DMRG level (energy, singlet-triplet gap energy).