| In this thesis I designed and implemented several algorithms aiming to increase effectiveness of classical and ab-initio molecular simulations. The developed algorithms and programs are based on information theory and machine learning theory. The first proposed algorithm tackles the problem of comprehensiveness of sampling and quality of classical force fields. In order to calculate the populations of conformational macrostates in a way that calculated populations can be compared directly to populations obtained via experimental methods, the new fuzzy clustering algorithm in collective variable space was developed. The proposed fuzzy clustering algorithm is robust and is capable to distinguish macrostates based only on data and without any prior macrostate definition. The second proposed algorithm also in a scope of classical molecular dynamics, utilizes information theory and targets both the quality of representation of free energy landscapes and speeding up of calculations of the rough free energy profiles. The algorithm uses adaptive mesh data reduction procedure, coupled with entropy based automatic adaptive bias potential generation. The later is driven only by a data set, i.e. there are no need of predefined or empirically derived input parameters. The data reduction step preserves the variability of data, but also creates parallel data structure, which can be implemented in any parallel scheme (MPI, OpenMP, Hybrid etc.). Thus the developed code is highly portable and can be used on various types of computers. Finally I address the problem of effective ab-initio molecular dynamics calculations (AIMD). It is well known fact, that the power and flexibility of AIMD comes at the price of a significant increase in computational overhead compared to empirical force field or semi-empirical approaches. That is why it is critical that AIMD calculations be designed in such a way that they can leverage large-scale parallel computing resources and increasingly popular accelerators such as graphical processing units (GPUs) with the greatest possible efficiency via the underlying computational algorithms. To address the issue, in this thesis I developed graphical processing unit (GPU) library for optimal calculation in density functional based molecular dynamics with localized basis set (discrete variable representations). The latter comprising the focus of the last chapter of the thesis. The GPU-DVR library was developed as separate module plugged to PINY parallel molecular simulation code which allows the simulation package to be used on various types of parallel architectures with many or without any number of GPU installed. |
