We reexamine some fundamental Quantum Monte Carlo (QMC) algorithms with the goal of making QMC more mainstream and efficient. Two major themes exist: (1) Make QMC faster and cheaper, and (2) Make QMC more robust and easier to use. A fast "on-the-fly" algorithm to extract uncorrelated estimators from serially correlated data on a huge network is presented, DDDA. A very efficient manager-worker algorithm for QMC parallelization is presented, QMC-MW. Reduced expense VMC optimization procedure is presented to better guess initial Jastrow parameter sets for hydrocarbons, GJ. I also examine the formation and decomposition of aminomethanol using a variety of methods including a test of the hydrocarbon GJ set on these oxygen- and nitrogen-containing systems. The QMC program suite QMcBeaver is available from the authors in its entirety while a user's and developer's manual is attached as supplementary material.
We reexamine some fundamental Quantum Monte Carlo (QMC) algorithms with the goal of making QMC more mainstream and efficient. Two major themes exist: (1) Make QMC faster and cheaper, and (2) Make QMC more robust and easier to use. A fast "on-the-fly" algorithm to extract uncorrelated estimators from serially correlated data on a huge network is presented, DDDA. A very efficient manager-worker algorithm for QMC parallelization is presented, QMC-MW. Reduced expense VMC optimization procedure is presented to better guess initial Jastrow parameter sets for hydrocarbons, GJ. I also examine the formation and decomposition of aminomethanol using a variety of methods including a test of the hydrocarbon GJ set on these oxygen- and nitrogen-containing systems. The QMC program suite QMcBeaver is available from the authors in its entirety while a user's and developer's manual is attached as supplementary material.
The exponential growth in computer power over the past few decades has been a huge boon to computational chemistry, physics, biology, and materials science. Now, a standard workstation or Linux cluster can calculate semi-quantitative properties of moderately sized systems. The next step in computational science is developing better algorithms which allow quantitative calculations of a system's properties. A relatively new class of algorithms, known collectively as Quantum Monte Carlo (QMC), has the potential to quantitatively calculate the properties of molecular systems. Furthermore, QMC scales as O(N3) or better. This makes possible very high-level calculations on systems that are too large to be examined using standard high-level methods. This thesis develops (1) an efficient algorithm for determining "on-the-fly" the statistical error in serially correlated data, (2) a manager-worker parallelization algorithm for QMC that allows calculations to run on heterogeneous parallel computers and computational grids, (3) a robust algorithm for optimizing Jastrow functions which have singularities for some parameter values, and (4) a proof-of-concept demonstrating that it is possible to find transferable parameter sets for large classes of compounds.
The goals of our research have been to: (i) expand the applicability of the quantum Monte Carlo (QMC) methods to larger molecular systems and check the reliability of traditional approaches, (ii) determine the impact of correlation energy on a variety of systems in which correlation is important, (iii) investigate new silicon clusters which may generate unique materials, (iv) determine the most stable C$\sb{10}$ and C$\sb{20}$ clusters, (v) achieve kcal/mol accuracy for molecular reactions, and (vi) develop the QMC methodology by improving the QMC trial functions. In order to attain these goals, we have unambiguosly determined the role of correlation energy in silicon and carbon systems, calculated binding energies, barrier heights, heats of formation, and electron affinities for selected molecular systems with high accuracy, and gauged the performance and predictability of more than ten standard methods. We have also shown that the utilization of natural orbitals leads to a significant improvement in the quality of the trial wavefunctions. We demonstrate that, through a combination of advances in both accuracy and scaling, QMC is a highly attractive alternative to the traditional methods, and for some cases it appears to be the only currently available approach to provide parameter-free predictions for energies of systems with more than 40 electrons.
A manager-worker-based parallelization algorithm for Quantum Monte Carlo (QMC-MW) is presented and compared with the pure iterative parallelization algorithm, which is in common use. The new manager-worker algorithm performs automatic load balancing, allowing it to perform near the theoretical maximal speed even on heterogeneous parallel computers. Furthermore, the new algorithm performs as well as the pure iterative algorithm on homogeneous parallel computers. When combined with the dynamic distributable decorrelation algorithm (DDDA) [Feldmann et al., J Comput Chem 28, 2309 (2007)], the new manager-worker algorithm allows QMC calculations to be terminated at a prespecified level of convergence rather than upon a prespecified number of steps (the common practice). This allows a guaranteed level of precision at the least cost. Additionally, we show (by both analytic derivation and experimental verification) that standard QMC implementations are not "perfectly parallel" as is often claimed.
Computing accurate yet efficient approximations to the solutions of the electronic Schrödinger equation has been a paramount challenge of computational chemistry for decades. Quantum Monte Carlo methods are a promising avenue of development as their core algorithm exhibits a number of favorable properties: it is highly parallel and scales favorably with the considered system size, with an accuracy that is limited only by the choice of the wave function Ansatz. The recently introduced machine-learned parametrizations of quantum Monte Carlo Ansätze rely on the efficiency of neural networks as universal function approximators to achieve state of the art accuracy on a variety of molecular systems. With interest in the field growing rapidly, there is a clear need for easy to use, modular, and extendable software libraries facilitating the development and adoption of this new class of methods. In this contribution, the DeepQMC program package is introduced, in an attempt to provide a common framework for future investigations by unifying many of the currently available deep-learning quantum Monte Carlo architectures. Furthermore, the manuscript provides a brief introduction to the methodology of variational quantum Monte Carlo in real space, highlights some technical challenges of optimizing neural network wave functions, and presents example black-box applications of the program package. We thereby intend to make this novel field accessible to a broader class of practitioners from both the quantum chemistry and the machine learning communities.
Computing accurate yet efficient approximations to the solutions of the electronic Schrödinger equation has been a paramount challenge of computational chemistry for decades. Quantum Monte Carlo methods are a promising avenue of development as their core algorithm exhibits a number of favorable properties: it is highly parallel and scales favorably with the considered system size, with an accuracy that is limited only by the choice of the wave function Ansatz. The recently introduced machine-learned parametrizations of quantum Monte Carlo Ansätze rely on the efficiency of neural networks as universal function approximators to achieve state of the art accuracy on a variety of molecular systems. With interest in the field growing rapidly, there is a clear need for easy to use, modular, and extendable software libraries facilitating the development and adoption of this new class of methods. In this contribution, the DeepQMC program package is introduced, in an attempt to provide a common framework for future investigations by unifying many of the currently available deep-learning quantum Monte Carlo architectures. Furthermore, the manuscript provides a brief introduction to the methodology of variational quantum Monte Carlo in real space, highlights some technical challenges of optimizing neural network wave functions, and presents example black-box applications of the program package. We thereby intend to make this novel field accessible to a broader class of practitioners from both the quantum chemistry and the machine learning communities.
Quantum Monte Carlo (QMC) is one of the most promising methods for solving quantum many-body problems. QMC combines known analytical properties of wave functions, results from other methods, and stochastic techniques, into a powerful tool for investigation of the electronic structure of real materials. Furthermore, the method has a favorable scaling ( O[N3] in the number of electrons), perfect scalability on parallel architectures, and can be applied to a wide range of systems. Our work has focused both on expanding the capabilities of QMC for calculation of quantities beyond energies, such as interatomic forces, and on application of the method to challenging problems in nanostructure materials research. For the calculation of interatomic forces we have implemented a finite difference correlated sampling method. The correlated sampling enables us to avoid the problem of statistical noise in evaluating the energy differences and allows us to obtain forces with an average 1% error. We have analyzed the effect of the method's main approximations on the estimate of the forces, both by theoretical arguments and by performing a variety of tests on small systems. Furthermore, we have used this newly developed method for estimating equilibrium geometries in a problem where other methods have failed. QMC enabled us to treat the electron correlation effects with high accuracy and to provide accurate predictions in a number of interesting applications. In particular, for carbon rings of C4N+2 stoichiometry we have resolved a long-standing problem regarding the competition between the aromaticity and second order Jahn-Teller effects. We have found that the dominant mechanism at small and intermediate sizes, which stabilizes the bond angle and bond length alternated geometries, is the secondorder Jahn-Teller effect while the aromatic isomer is always found to be a transition state. These high accuracy calculations which have involved multi-reference wave functions, extensive variational optimizations, and the use of massively parallel platforms demonstrate clearly the unique capabilities of the QMC approach.