Prof. Dr. rer. nat. Karsten Reuter

Courses and Dates

Offered Bachelor’s or Master’s Theses Topics

Bringing Solvent Structure to Implicit Solvation

With solvent effects being central to several scientific disciplines like biology or (electro-)chemistry, computationally affordable methods are crucial to accurately simulate solvated systems. Implicit solvation models are widely used due to their simplicity and practical applicability in these kind of many-body problems. Here only the solute is explicitly modeled while the interaction with the solvent is coarse-grained into a continuum dielectric response, instead of considering individual solvent molecules. However, simply using the solvent medium’s experimental bulk dielectric constant neglects the fact that the behaviour of the solvent differs, sometimes drastically, in the vicinity of the solute compared to the bulk. Existing corrections to account for these entropic effects fail for charged systems and lack scientific foundation. These deficits of correctly accounting for multiple conformations in solution lead to great difficulties when calculating dissociation
constants (pKa values) or modeling other electrochemical processes using implicit solvation models.

The goal of this M.Sc. project is to introduce a better description that accounts for the orientational behaviour of solvent modecules, based on findings from previous extensive Molecular Dynamics simulations. The student will learn how to perform DFT calculations using implicit solvation and, in a second step, will compute the interaction of the interfacial water structure with the electric field obtained from these simulations. Basic knowledge of UNIX based operating systems and some experience with programming (and/or scripting) languages (preferably Python) is desirable, but not mandatory.
1

suitable as

• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter

Deep Neural Network methods to generate new Organic Semiconductors

This Thesis is based on Deep Neural Networks (DNN) applied to generate new organic semiconductor (OS) molecules. Deep learning models project input data through several layers of nonlinearity and learn different levels of abstraction generating a link between the inputs and specific target outputs (or labels). Recently, there has been a fundamental work about using DNN algorithms for synthesising in-silico new OS [1] grounded on the seminal work by Gomez-Bombarelli et al. [2] and Kusner et al. [3] on the Grammar Variational AutoEncoders (GVAE) method. The thesis is focused on implementing a GVAE and further developing the method to include bigger molecules. The DNN will be applied to some realistic OS in order to evaluate the efficiency and accuracy in generating new molecules.

[1] P. B. Jørgensen et al., J. Chem. Phys., 148, 241735 (2018)

[2] R. Gómez-Bombarelli et al., Nat. Mat., 15, 1120-1128 (2016)

[3] M. J. Kusner, B. Paige, J. M. Hernández-Lobato, Grammar Variational Autoencoder, arXiv:1703.01925v1 [stat.ML] (2017)

suitable as

• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter

Machine learning for computational design of nanostructured catalysts

The computational design of new materials for heterogeneous catalysis relies on methods that can accurately predict adsorption energies of reactants and intermediates at low computational cost. Quantum-mechanical calculations based on the Density Functional Theory (DFT) are typically used for surface reactivity studies, but their cost significantly limits the number of compounds that can be evaluated as candidates for catalysing each targeted reaction. Scaling relations between adsorption energies of similar adsorbates and machine-learning approaches have been recently developed to circumvent this computational burden, but their application is currently limited to extended close-packed metal surfaces. Real catalysts are typically formed by supported metal nanoparticles exposing sites at different facets, edges, and corners. Dealing only with extended surfaces thus limits the predictive potential of computational catalyst design.

 

The goal of this M.Sc. project is to extend existing approaches for predicting adsorption energies to the description of undercoordinated sites typically found on metal nanoparticles. The candidate should be interested in electronic structure and machine learning methods, as well as in surface chemistry and catalysis. The student will learn how to perform DFT-based calculations and to use and asses the results of different machine learning methods. Basic knowledge of UNIX based operating systems and some experience with programming (and/or scripting) languages (preferably Python) is desirable, but not required.

 

Questions about the project can be directed to mie.andersen@ch.tum.de and albert.bruix@ch.tum.de. 

 

suitable as

• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter

Origin of the open circuit voltage in ternary organic solar cells, a study based on a semi-empirical Monte Carlo simulation approach

Organic photovoltaic (OPV) technology has seen a rapid evolution over the last decade. The unique selling points of OPVs, such as excellent light harvesting capability, freedom of form, color and transparency, environmental friendliness, easy scalability and lower manufacturing costs based on roll-to-roll printing methods, position this technology for the mobile power market, and this most properly reflects the state of the art in commercialization. An important milestone towards OPV commercialization has been surpassed by reaching a power conversion efficiency (PCE) of up to 17%. The main limitation in OPVs is due to the intrinsic narrow absorption window (~100-200 nm) of polymers compared to inorganic semiconductors such as Si, which makes it challenging to fully cover the solar spectrum with a single junction device. To overcome the absorption limitation, ternary blend organic solar cells represent one of the dominant strategies that has been explored in the last decade. The outstanding advantage of ternary blends consists of maintaining the simplicity of the processing conditions used for single active layer devices. Interestingly, the open circuit voltage (Voc) is tunable depending on the loading concentration of the ternary compound in both ternary systems compromised of i) two donors and one accepter as well as ii) one donor – two accepters blends.

The origin of the composition-tunability of Voc and the optimal electronic correlations between the components remains as an open question in OPV field. Obviously, it is very challenging to develop a complete model for Voc in ternary systems mainly due to the fact that this model should take all the electronic interactions between the components into account and consider their role in the photogeneration. Moreover, the model should be compositional dependent and accounts for the different microstructures and transport mechanisms operating simultaneously in the system. To this end, a semi-empirical method is required to link the electrical characteristics of ternaries to their morphological properties and draw a comprehensive picture from the morphology models reported in literature as the origin of Voc changes in ternary systems.

This project will focus on solving the aforementioned key challenge by employing a semi-emprical method based on kinetic Monte Carlo (kMC). An existing lattice kinetic Monte Carlo model, previously applied to the modeling of binary polymer:fullerrene solar cells, will be extended to treat ternary blends. This will allow us to correlate nano-morphological features with measured optical properties and obtained Voc, a crucial capability for the design of optimized, high performance ternary solar cells. 

suitable as

• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter

Verzerrungen in LSPS Kristallen: Einblicke durch maschinelles Lernen

In 2011, a new solid lithium electrolyte was reported, featuring liquid-like Li-ion conduction in a crystalline solid matrix. The ultrafast room temperature transport of tetragonal Li10GeP2S12 (LGPS) with a conductivity of several mS/cm exceeds the values of most crystalline Li conductors by one order of magnitude. Accordingly, there has been a strong search in realizing LGPS-type materials based on the homologous elements Si and Sn. LXPS (X=Ge,Sn,Si) electrolytes appear in a tetragonal and orthorhombic modification. While the orthorhombic phase is an undesired inpurity in the LGPS electrolyte, orthorhombic LSPS leads to an enhancement of the conductivity. A possible explanation of the increased conductivity,
is an interplay of orthorombic and tetragonal LSPS, which distorts tetragonal LSPS. This distortion may lead to a favorable opening of Li-channels. The aim of this work is to investigate the effect of distortion on the materials conductivity for tetragonal LSPS. To do so a machine-learned force-field will be provided. The project will start with the construction of distorted LSPS structures and will then focus on Molecular Dynamics (MD) calculations for appropriate LSPS ensembles at finite temperatures. Finally, the conductivities will be calculated via the Nernst-Einstein relation.

suitable as

• Bachelor’s Thesis Physics
• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter

Wie verrauscht sind eigentlich Elektronenstrukturtheorie-Daten?

The combination of quantum mechanical calculations and machine-learning techniques has led to massive advances in the high-throughput computational screening of materials and molecules. Here, it is frequently stated that training data obtained from quantum mechanical simulations (e.g. using density functional theory) is 'noise-free', because the calculations can be numerically converged to very high precision. However, this assumes that the self-consistent algorithm used to calculate the Kohn-Sham determinant always converges to the correct ground-state solution, which is by no means guaranteed. The goal of this project is to quantify the noise introduced into DFT data by different choices of intital guess and convergence acceleration. To this end, the true ground-state solutions for a representative set of molecules will be determined using wavefunction stability analysis. This benchmark will allow the development of an effective 'black-box' method for obtaining well-converged training data. Furthermore, the effect of noise on quantum mechanical machine-learning applications will be studied.

The candidate should be interested learning the fundamentals of density functional theory and supervised machine learning. The project will require the development of automated workflows for high-throughput electronic structure calculations and data analysis. Experience with a scripting language (e.g. Python) and UNIX operating systems is advantageous but not mandatory.

Questions about the project can be directed to johannes.margraf@ch.tum.de.

suitable as

• Bachelor’s Thesis Physics
• Master’s Thesis Condensed Matter Physics
• Master’s Thesis Applied and Engineering Physics

Supervisor: Karsten Reuter