This proposal aims to develop parallel linear scaling algorithms for density functional theory (DFT) calculations. DFT calculations are the most widely used electronic structure methods, because at present they offer the best compromise between computational cost (typically between linear and cubic scaling) and accuracy (typically a few kcal/mol for chemical reactions). DFT calculations are widely used for exploring chemical reaction mechanisms, and there is strong demand to increase the size of treatable systems at the current limit of feasibility. The opportunity is to leverage the greatly increased power and greatly reduced cost of workstation/PC clusters in combination with linear scaling algorithms that replace the diagonalization step. This step is at present most challenging in very large-scale DFT calculations, and is also least satisfactorily parallelized in conventional DFT codes. New linear scaling algorithms with increased efficiency will be sought, and customized tools for performing sparse linear algebra in parallel will be developed. The result will be the first widely distributed electronic structure program with this twin capability of scalability with respect to number of processors (via efficient parallelization) and scalability with respect to problem size (via true linear scaling algorithms).
Liang, WanZhen; Head-Gordon, Martin (2004) An exact reformulation of the diagonalization step in electronic structure calculations as a set of second order nonlinear equations. J Chem Phys 120:10379-84 |