The Hauser Group
 

Research

 
  Our research spans over applications of quantum chemistry to a broad selection of problems (cluster physics, catalysis, gas separation and storage, porous materials, chiral separation, molecular sieving), fundamental concepts of molecular spectroscopy, and method development with a special focus on machine learning implementations. Currently, we are particularly interested in the following projects:

Surface chemistry and metal cluster physics

The properties of metals are strongly dependent on the actual particle size. Well defined nanoparticles, with atoms organized in typical crystalline facets, behave differently from monodispersed subnanometer clusters with their many undercoordinated sites. Additional complexity is introduced due to differences in shape and the interaction with the support. Currently, we are investigating the intermetallic interactions in sub-nanometer sized, mixed-metallic particles on their chemical and physical properties.

Chiral separation

In a recent publication we suggested a new concept for the separation of chiral molecules via single-atom-thick membranes. It is based on molecular identification upon pore entrance and has the potential to exceed common techniques such as gas chromatography or high performance liquid chromatography both in terms of efficiency and cost, and might also help to reduce the environmental impact of chiral separations in the chemical and pharmaceutical industries. Currently, we are aiming at a further development of the proposed concept via a combination of electronic structure theory and molecular dynamics simulations. Our last article on that matter just came out recently.

Quantum technology and ultracold chemistry

In a tight collaboration within the MOLIM framework ("Molecules in motion", a EU COST project) with María Pilar de Lara-Castells from the CSIC in Madrid we study the effects of a superfluid helium environment on large molecules such as fullerenes and carbon nanotubes. We are interested in the effects of the helium bath on intermolecular interactions, e.g. between C60 and an alkali metal atom. Can an exothermal reaction be avoided if one dopant resides inside a He droplet while the other is "floating" on its surface? Another question we ask is how does the wetting behavior of superfluid helium differ from the classic picture, e.g. "How is the sinking an open carbon nanotube in superfluid helium different from sinking a tin can in the sea?" These questions might also be relevant in the context of quantum simulation and quantum information, where charged particles are kept floating on surfaces of liquid helium.

Machine learning approaches in computational chemistry

Currently, we are undertaking several attempts to investigate the applicability of machine learning approaches to various fields of computational chemistry. Most of these investigations feed into a long term goal of our current national projects (FWF projects P 29893-N36 and PIR 8-N34), which is to perform structural studies on mixed-metallic nanoparticles, systems with sizes that make a meaningful exploration of potential energy surfaces extremely time consuming or even impossible with standard approaches. Related to this, or rather an extension of it, is the attempt to predict chemical and optical properties of metallic nanoparticles in the nanometer-range for heterogeneous catalysis, photocatalysis or optical applications. At the moment, special focus is set on the following four ideas.

Method improvements to multi-layered feed-forward NNs

Behler-Parinello-type neural networks have proven to be a useful tool for certain tasks in quantum chemistry. However, the large amount of data points needed for successful training limits the benefit over direct methods. Also, training data generation is often challenging, in particular for metallic systems with troublesome convergence behavior, and especially in regions far from from the equilibrium geometry. Therefore, we attempt to arrange these networks into a physically motivated form which enforces a correct behavior for short inter-atomic distances by construction.

Acceleration of SCF algorithms via improved guesses

Suitable initial guesses for the self consistent field (SCF) iteration algorithms of Hartree-Fock implementations have seen many improvements in the last decades. However, enforcing convergence e.g. for large metallic systems is still a problematic case, and automatic black box-type initial guesses often apply only to a limited selection of standard basis sets. Hückel-based methods are very effective for organic molecules, while metallic systems are preferably treated via the conceptually simple superposition of atomic densities. This is the starting point for our machine learning ansatz, which uses basis set and molecular structure information in the form of the overlap matrix S to predict the matrix elements of the density matrix.

Transition state search accelerated with machine learning algorithms

Well established methods such as nudged elastic band calculations, growing or freezing string methods and eigenvector following techniques are currently used for the calculation of reaction pathways. Yet, they all typically operate directly on the potential energy surface to be explored, which makes them rather time consuming, especially in the case of large systems with complicated high-dimensional surfaces. Two promising approaches, the application of neuronal networks and Gaussian process regression have been suggested recently. We investigate two alternative kernel-based machine learning strategies, namely support vector machine and ridged regression. A big advantage of all four algorithms in comparison to standard evaluations of the reaction pathway is the additional information gained about the local shape of the PES. Furthermore, Gaussian process regression also offers error estimates, which makes this technique particularly interesting.

Embedded atom model potentials for metallic nanoparticles derived from NNs

Reliable interatomic potentials are crucial for the simulation of large systems. Machine learning approaches are typically applied to purely mathematical potential forms for the sake of generality, most notably in the atomic neural network approach suggested by Behler and Parinello. Ironically, these methods often fail at the generalization to other systems as they learn only system-specific, lacking any physical understanding such as e.g. a correct asymptotic behavior. On the other hand, physically meaningful force field models are not flexible enough to reproduce ab initio surfaces. Therefore, we aim to bridge between these two extremes by also modeling physical concepts within a neural network.

Porous membranes and graphene sheets

The chemical inertness and two-dimensionality of porous graphene makes it an interesting material for applications of gas separation and storage. Its permeability for a certain gas species is mainly determined by the pore size, making it the essential parameter for separation adjustments. However, the propagation of a molecule through pores with diameters of a few angstrom needs to be treated quantum-mechanically. This becamce particularly obvious in our studies on the separation of helium isotopes based on counteracting effects of quantum tunneling and zero-point energy.

Molecular spectroscopy

A good part of our research time is also spent on the assignment and interpretation of molecular spectra measured by our experimental collaborators. A big advantage of the exotic experimental technique of the Ernst group (Helium Nanodroplet Isolation Spectroscopy or HENDI) is that it gives access to spectra of rather exotic and very weakly bound species such as van de Waals-bound metal clusters in high spin states. We support their studies with ab initio calculations of ground- and electronically excited states and provide model Hamiltonian descriptions and computations for cases of spin-orbit and non-adiabatic coupling.

Collaborations

Prof. Wolfgang E. Ernst

Our currently most active collaboration is set up with the group of Wolfgang Ernst, who also happens to be the head of our institute. There is great synergy between our theoretical interest in the physical and chemical properties of sub-nanometer sized mixed-metallic particles and the experimentalists' ability to create these particles via a complicated heliumdroplet deposition procedure. We further support his group with the theoretical assignment of measured laser-induced fluorescence spectra, which often involves the calculation of highly excited electronic states including non-adiabatic effects, spin-orbit coupling and in rare cases even information on hyperfine structure details.

Prof. María Pilar de Lara-Castells

With the group of María Pilar de Lara-Castells from the CSIC in Madrid we collaborate on the theoretical treatment of molecular systems with pronounced quantum phenomena, typically occuring at very low temperatures. A lot of our recent reserach focussed on molecules in a superfluid helium environment, studying the interplay of dispersion forces and quantum kinetic energy.

Prof. Peter Schwerdtfeger

The collaboration with the group of Peter Schwerdtfeger in New Zealand has lasted now for several years. Originally, it focussed on the reduction of methane emissions. This gas, which is released by farm animals as well as landfills, crops and coal mines, is known to be much more effective in global warming than carbon dioxide.

Our approach was to comb through the huge variety of newly developed, highly porous nanomaterials and determine the most promising candidates for a cost-effective in situ separation of CH4 from air at room conditions.

We could show that porous graphene sheets with neatly adjusted pore diameters have the potential to separate methane from other gases. This discovery triggered a whole series of publications in the field of gas separation, and even led to a novel concept for the separation of bosonic from fermionic helium, a separation based on pure quantum effects!

Prof. Martin Head-Gordon

Martin Head-Gordon, professor of theoretial chemistry at UC Berkeley, is world reknown for his contributions to method development in the field of quantum chemistry. Besides, he is also one of the founders and the scientific advisor of the Q-Chem program package. As a part of the Q-Chem developer community, we are using, debugging and extending the package, always trying to push its current limitations.

A current side project of our group is the development of the Python module pyQChem, which aims at improving the usability of Q-Chem (simplified parsing, input batch file design and job programming, extensions for the calculation of thermodynamics properties).

Scientifically, we are currently involved with recent work of the Head-Gordon group on Energy Decomposition Analayis (EDA) and its application to small metal clusters in order to study their catalytic properties.

 
Maintained by Andreas Hauser | Last updated June 2018 | Impressum