Investigating quantum Monte Carlo methods in Slater determinant bases
William Andrew Vigor, Michael Bearpark, Alexander Thom, Engineering And Physical Sciences Research Council
In this thesis we investigate the recently developed Full Configuration Interaction Quantum Monte Carlo (FCIQMC) method. This method, unlike the traditional methods in Quantum Monte Carlo (QMC), doesn't require the use of the uncontrolled fixed node approximation and so potentially it can yield far more accurate results. This means that it can be thought of as a hybrid between the methods used by quantum chemists and QMC, and thus has spawned a new field of stochastic quantum chemistry. The
... described in this thesis can be split into three distinct but interrelated parts. We begin with an investigation of the underlying FCIQMC stochastic process. We show that FCIQMC is an example of Markov Chain Monte Carlo. This means we can compute a stochastic matrix from which all details about the Monte-Carlo simulation can be obtained. Unfortunately the size of the space scales unfavourably as a function of system size meaning that we have only managed to compute the matrix for the smallest interesting two determinant system. We then use these results to quantify population control bias in FCIQMC for a two determinant system supplementing these analytical results with empirical results to investigate more realistic systems. We then attempt to quantify the efficiency of the FCIQMC algorithms, defining a measure of efficiency. After this we investigate the dependence of our measure on the system size. The error bar of the most efficient FCIQMC algorithm will decay fastest as a function of computer time. We then draw conclusions about the applicability of the FCIQMC method. Finally we describe an implementation of FCIQMC on a novel data flow computer architecture. In our implementation we made a modification to the FCIQMC algorithm to fit the data flow paradigm. We investigate the efficiency of the modified FCIQMC algorithm.