Hole A Particle Corresponding To The Mrci Program And Its Parallel

Multi-reference configuration interaction with single and double excitation (MRCISD) is regarded as a reliable way to calculate the electronic correlation energy. MRCISD have been used popularly in works such as excited state, chemical reaction and potential energy surface (PES) etc. The bottleneck that restricts the usage of this method is the efficiency of MRCISD. To obtain an accurate result, it is necessary to choose large basis sets and a large number of reference functions, which lead to a rapid increase of the CI space and become computational intractable. Furthermore, to calculate the PES, thousands of energy points are needed to be deal with, hence, considerable resources and run time are needed to perform MRCISD calculations. An effective MRCISD program is highly desired.In chapter one, we introduce the concept of the electronic correlation energy and some methods that are used to calculate the correlation energy.Because our MRCISD program is based on Graphic Unitary Group Approach (GUGA), in chapter 2 we briefly review some concrete problems in applying GUGA to CI calculation. The many-body electronic Hamiltonian is realized as a polynomial in the U(n) generators with the one- and two-electron integrals as coefficients. The methods for the computation of the matrix representation of the Hamiltonian over Gelfand states are discussed.In chapter 3, we discuss a new MRCISD algorithm which is based on the hole-particle symmetry approach. At first, a new classification of spin adapted MRCISD states using hole-particle symmetry is suggested. Second, the formulas and values of 244 hole loop shapes are derived according as the rule of transform from external space loop shapes to hole loop shapes. Finally, a new CI algorithm is presented and coded, which is efficient and disk storage saving indicated from some given examples.In chapter 4, the parallel implementation of multi-reference configuration interaction program is described. MPI and Global Array toolkit are used as the parallel libraries. Using theshared-memory model which is provided by GA, we design the task parallelism algorithm for HC step. The task is defined as the entire works to be done to update two segments of the CI vector. The segmentation of the CI vector is based on the DRTs in active and hole space. The algorithm referred as distributed task pools is used as the dynamic load balancing algorithm in HC step. The static load balancing is used in sub-space operation. Since sub-space operation is IO intensive, we use the distributed file storage scheme.In chapter 5, we introduce an approximate CI algorithm---Doubly contracted CI (DCCI). Due to the similarity between the hole space and the external space in the DRT structure, the contraction procedure used in the externally contracted CI can also be used in the hole space. We discuss the detail of the DCCI method and give some benchmark examples. The loss of correlation energy due to DCCI is in general larger than external or internal contraction, but the examples show the loss of relative correlation energy is small and a satisfactory accuracy can be expected. This method is efficient and further research is needed.