Overview

Catalytic processes are implicated in everyday applications related to energy production, agriculture, medicine, and various other fields. Many catalytic processes are sufficiently complex, meaning experimental measurements cannot be used to accurately quantify various details of the reaction pathways and kinetics. Sophisticated computational models of chemical catalysis are therefore necessary to both understand and predict the outcomes of these reactions, and to inform future catalyst development. One such example of catalysis can be found in the reduction of carbon dioxide into useful two-carbon chemicals using copper electrodes as the catalyst. This system is of interest as a potential target for efficient production of energy and chemical feedstocks which also provide a viable pathway to carbon recycling. We investigate the myriad scientific difficulties posed by this system while developing computational tools which can also be transferred to the study other catalytic systems of chemical and societal interest.

Challenges

Physical Sciences Challenges

The chemical reactions involved in electrocatlytic systems pose diverse challenges for physical modeling. Various details of a catalytic cell—solvation, mass transport, pH dependence, space-charge fields, specific ion effects—play non-trivial roles in determining reaction rates and pathways; all of which (and more) must be included in predictive models. Accurate determination of the energetics of chemical reactions and binding is also of fundamental importance. These challenges will be addressed through the use of existing hybrid density functionals and through the development of new functional tuned specifically to treat extended metallic systems. At the same time, density functional embedding theoreatical approaches will be used to achieve chemical accuracy for the catalytic centers. Explicit treatment of molecular solvation and the electrode-electrolyte interface will address limitations of the commonly used continuum representations of solvation. Since electrocatalytic systems do not operate under equilibrium conditions, non-equilibrium molecular dynamics will also be utilized in the study reaction pathways. Finally, nuclear quantum effects, zero point energies, and bond anharmonicity play important, albeit subtle, roles in determining various physical properties including electron density and bonding properties. These will be addressed through the application and development of density functional theories which include quantum effects for the pertinent nuclei.

Computer Science Challenges

Modern electronic structure theories and molecular dynamics place stringent demands on computational resources. Efficient implementations of these physical models require: scalable algorithms for sparse and dense linear algebra, iterative eigensolvers, quadrature on various types of grids, and management of large amounts of data. Progress has been made on each of these fronts, and we will meet remaining challenges through better leveraging of modern supercomputer architectures. This will be accomplished in part through more scalable approaches to hybrid programming models. Additional work will also address the difficulties in efficient data management and transfer in highly parallelized systems. Algorithm optimizations will also employ elastic resource management in order to achieve solutions in an energy sufficient manner. While this work will be completed in the context of computational chemistry frameworks, the subsequent developments in high-performance computing will be broadly transferable within the DOE scientific community.

Applied mathematics challenges

Cutting edge electronic structure theories require mathematically sophisticated implementations in order to achieve computational efficiency. In particular, current approaches with hybrid density functionals and standard basis sets are intractable when applied to systems of interest in electrocatalysis. These challenges will be addressed through development of adaptively compressed exchange, compressed representations of density matrices, as well as interpolative separable density fitting. Our approach will dedicate particular focus on calculations which make use of large basis sets; necessary to achieve chemical accuracy for the systems of present interest. Further work will use preconditioning to minimize the number of iterations necessary to achieve self-consistency in Kohn-Sham eigensystems. The numerics associated with the various embedding theory approaches will be improved through the further development of adaptive local basis set approaches.

Publications

Materials

Software

Libtensor
Tensor algebra library for computational chemistry. The library is a set of performance linear tensor algebra routines for large tensors found in post-Hartree-Fock methods. The routines are able to dynamically balance the use of physical memory (RAM) and disk and perform computations on multiple CPUs or CPU cores in a shared-memory environment.
Download Libtensor 2.2 (BSD License)
If used, please cite J. Comput. Chem. 2013, 34, 2293.

Q-Chem
Q-Chem is a comprehensive ab initio quantum chemistry package for accurate predictions of molecular structures, reactivities, and vibrational, electronic and NMR spectra. With Q-Chem as the development platform for the new electronic structure codes, we have developed and implemented the following methods:
• RAS-SF : Variational method for describing excited states and strongly correlated molecules
• SF-CAS(h,p), SF-CAS(S) : Perturbative methods for describing excited states and strongly correlated molecules, including the possibility for including dynamical correlation.
• SF-NOCI : Variational method which uses non-orthogonal orbitals to achieve a more compact wavefunction, which still includes orbital relaxation.
• Complex Basis-Functions : An open-ended and efficient implementation of complex basis functions for describing low energy shape resonances in unbound systems.

STRUMPACK
STRUctured Matrices PACKages - is a package for computations with sparse and dense structured matrices, i.e., matrices that exhibit some kind of low-rank property. STRUMPACK uses Hierarchical Semi-Separable representations (HSS) and Randomized Sampling techniques.