Item

Abstract

Two essential issues of semiempirical quantum chemical methods are addressed in this dissertation:performance and accuracy. First, semiempirical program
code was developed for heterogeneous high-performance CPU-GPU computing platforms. In systematic test calculations on large molecules, the overall performance could be improved by one order of magnitude, which is unattainable on CPU-only parallel computers due to intrinsic constraints of the hardware architecture. Second, both local and global optimization algorithms for the parameters of semiempirical methods were implemented from scratch. The efficiency of parameterization was increased by taking advantage of coarse-grained parallelism on symmetric multiprocessors, which enables more comprehensive explorations of parameter space. This was demonstrated by reparameterization of OM2 and OM3 using dispersion corrections and by specific parameterizations for an enzymatic reaction, the hydride transfer catalyzed by dihydrofolate reductase, and for hydrogen bonding and proton transfer in water. The optimized CPU-GPU code was used in asystematic benchmark with full geometry optimization for a set of 28 proteins using 10 different semiempirical quantumchemical methods. These extensive computations unveiled some limitations of the currently available semiempirical methods that need to be addressed in future work.