# Seminar Series

This seminar series is open to interested researcher and aims to foster the discussion of new theories and novel methods related to electronic structure theory, machine learning, molecular dynamics, and structure exploration.

If you want to join a seminar, please** register below**, we will then send you a link via e-Mail. We are looking forward to fruitful discussions!

**Summer term 2022**

**April 29, 2022, 10.00 am GMT**

**Electrochemistry and electrocatalysis from first principles**

**by Dr. Marko M. Melander**

###### Department of Chemistry, University of Jyväskylä, Finland

Electrochemistry enables efficient interconversion of electric and chemical energy which can be harnessed in numerous electrocatalytic reactions such as hydrogen evolution and oxidation, oxygen reduction, and CO_{2} reduction reactions. At the heart of all electrochemical and electrocatalytic lies the ability to control reaction thermodynamics and kinetics at solid-liquid interfaces through the application of an external electrode potential. While experiments are readily performed under such conditions, theory and simulation of electrochemistry under conditions mimicking experiments is extremely difficult requiring the development and application of advanced theoretical and modelling methods.

In this seminar, I will shortly present the basic concepts of electrochemistry and electrocatalysis, and how they can be modelled using first-principles methods. I will discuss some recently developed theoretical and computational methods for simulating electrocatalytic thermodynamics and kinetics under constant potential conditions within the grand canonical ensemble (GCE) approach. The theoretical framework underlying the GCE approach and its combination with density functional theoretical (DFT) will be discussed briefly. After this, some recent applications of GCE-DFT on electrocatalytic systems are presented. Overall, I will present the theory and methods for constant potential simulations at the GCE-DFT level, and how such simulations can be used to understand electrochemistry and electrocatalysis from the atomic level.

###### Register hereLink opens in a new window

**The event will be held in-person at Warwick in C521 (C block 5th floor seminar room).**

**May 6, 2022, 10.00 am GMT**

**Black-box algorithms and robust error control for density-functional theory**

**by Dr. Michael F. Herbst**

###### Applied and Computational Mathematics, RWTH Aachen University, Germany

Systematic first-principle calculations on thousands to millions of compounds have become an established tool in materials modelling. In this setting automation and reliability of simulations are of elevated importance, which translates to a requirement for numerical methods to be parameter-free and above all robust. At the same time not all calculations in such a screening workflow are required at equal accuracy. A promising outlook of the mathematical study of simulation errors is therefore not only reduce failure rates, but also to identify where approximate computing can be safely employed for gaining extra efficiency in simulations.

Motivated by this prospect this talk focuses on Kohn-Sham density-functional theory (DFT), the most widespread first-principle simulation method in the field. First we discuss the non-linear eigenvalue problem underlying DFT and propose two improvements to the standard self-consistent field (SCF) procedure commonly used for solving it: a preconditioner based on the local density of states as well as a line-search technique based on an approximate model for the DFT energy. Both methods are inspired from a mathematical analysis of the SCF problem, free of user-chosen parameters and robustly applicable to a wide range of systems. With a seamless integration into standard acceleration techniques (such as Anderson) these methods are moreover almost as fast or even faster than standard SCF approaches.

In the second part we discuss recent efforts to integrate solid-state DFT simulations with algorithmic differentiation. These techniques enable to automatically compute arbitrary derivatives of output quantities with respect to parameters of the model (DFT functional, pseudopotential etc.), i.e. without needing to implement such derivatives by hand. This is illustrated by exemplary computations tracking the sensitivity of DFT lattice constants on the chosen DFT functional.

###### Register hereLink opens in a new window

**The event will be held in-person at Warwick in C521 (C block 5th floor seminar room).**

**Spring term 2021/22**

**March 25, 2022, 10.00 am GMT**

**Electronic structure meets machine learning**

**by Dr. Jan Hermann**

###### Freie Universität Berlin, Kaiserswerther Str. 16-18, 14195 Berlin, Germany

###### Register hereLink opens in a new window

**The event will be held in-person at Warwick in C521 (C block 5th floor seminar room).**

**Winter term 2021/22**

**January 14, 2022, 1.00 pm GMT**

**Tensor-valued atomic cluster expansion for inference of dynamical systems**

**by Ass.-Prof. Dr. Matthias Sachs**

###### School of Mathematics, University of Birmingham, United Kingdom

[1] M. Bachmayr *et al.*, “Atomic Cluster Expansion: Completeness, Efficiency and Stability,” Nov. 2019.

[2] R. Drautz, “Atomic cluster expansion for accurate and transferable interatomic potentials,” *Phys. Rev. B*, vol. 99, no. 1, p. 014104, 2018.

###### Register hereLink opens in a new window

**Autumn term 2021**

**October 8, 2021, 1.00 pm BST**

**Charge Transfer Mechanisms & Selectivity in Heterogeneous**

**Electrocatalysis **

**by Dr. Vanessa J. Bukas***

###### Theory Department, Fritz-Haber-Institute of the Max-Planck-Society,

Faradayweg 4-6, 14195, Berlin, Germany

Heterogeneous electrocatalysis holds promise for the continued production of fuels and chemicals in a sustainable, fossil-free future. Despite substantial progress over the past decade, however, there is still an urgent need for improved electrocatalysts before clean energy technologies can be considered viable for widespread application. A key challenge in meeting this goal will be to better control catalyst selectivity, especially toward high-value products such as e.g. long hydrocarbons, oxygenates, and ammonia. Such efforts are presently hindered by our limited fundamental understanding of electrocatalytic selectivity in general. Even for extremely well-defined experimental systems, identity of the favored product is often unexplained and cannot be rationalized by popular thermodynamic models of concerted proton-electron transfers (CPETs) that are commonly used in computational electrocatalysis.

This talk will challenge prevalent mechanistic assumptions and discuss microscopic reaction mechanisms under conditions of limiting electron or proton availability, as a means to selectively suppress specific CPET steps or reaction routes. The idea will be showcased for a handful of systems where delayed charge transfer is suggested as a determining factor to the resulting product selectivity. Atomistic models, based mainly on density functional theory, will further be used to show that such mechanisms can explain key experimental findings and even be exploited to systematically steer product selectivity in technologically important electrochemical processes. This suggests departure from the common-practice approach of predicting activity trends based on thermodynamic descriptors of specific charge-neutral surface intermediates and underlines interfacial charge transfer kinetics as a promising research direction that may help gain better control over electrocatalytic selectivity.

###### Register hereLink opens in a new window

**October 29, 2021, 10.00 am BST**

**Ehrenfest TD-DFTB and its applications in transient absorption and impulsive Raman spectroscopies**

**by Dr. Franco Bonafé**

###### Max Planck Institute for the Structure and Dynamics of Matter, 22761 Hamburg, Germany

Several scientific and technological fields demand fast and efficient computational tools to understand the ultrafast quantum dynamics after photoexcitation of relatively large matter systems. Such tools should also be able to account for nuclear motion, as in many situations coupling of electronic excited states with molecular vibrations can be crucial to describe these processes. In this seminar I will describe an implementation of semiclassical Ehrenfest dynamics within time-dependent density functional tight-binding (TD-DFTB) in the free, open-source DFTB+ package. I will show some successful applications of this tool to describe light-induced impulsive vibrations in metal nanoparticles and molecules. Finally, I will explain how this method can be used to simulate transient absorption spectra and impulsive Raman spectra.

###### Register hereLink opens in a new window

**November 11, 2021, 10.00 am BST**

**Setting benchmarks for theoretical models of gas-surface scattering**

**by Dr. Helen Chadwick**

###### Department of Chemistry, College of Science, Swansea University, Swansea, UK

Several factors can affect the outcome of a collision of a gas phase molecule with a surface, from the velocity of the incoming molecule to the internal energy states of that molecule. The most stringent tests of the accuracy of theoretical models of these collisions are arguably provided by state-resolved experiments, which remove the averaging over the many degrees of freedom that influence the gas-surface interaction. One particularly challenging quantum state to control is the rotational orientation projection (m_{J}) state, especially for ground state molecules. Here, a new technique will be presented [1], which allows the rotational orientation of ground state molecules to be prepared, manipulated and controlled using a series of homogeneous and inhomogeneous magnetic fields. The results that have been obtained using this method for the scattering of H_{2} from LiF will be shown [2]. Analysis of the data shows that scattering from a LiF crystal can in itself produce a rotationally oriented H_{2} beam, as predicted theoretically over 20 years ago [3]. In addition, we can determine both the amplitude and the phase change of the wavefunction as the molecules scatter from the surface, which provides a further, more stringent test of current state of the art theory, and will help to develop and benchmark theoretical models.

**References**

[1] O. Godsi, G. Corem, Y. Alkoby, J.T. Cantin, R. V Krems, M.F. Somers, J. Meyer, G.J. Kroes, T. Maniv, and G. Alexandrowicz, Nat. Commun. **8**, 15357 (2017).

[2] Y. Alkoby, H. Chadwick, O. Godsi, H. Labiad, M. Bergin, J.T. Cantin, I. Litvin, T. Maniv, and G. Alexandrowicz, Nat. Commun. **11**, 3110 (2020).

[3] E. Pijper and G.J. Kroes, Phys. Rev. Lett. **80**, 488 (1998).

###### Register hereLink opens in a new window

**Summer term 2021**

**April 30, 2021, 1.00 pm BST**

**Tunneling and Zero-Point Energy Effects in Multidimensional Hydrogen Transfer Reactions**

**by Dr. Yair Litman**

###### Max Planck Institute for the Structure and Dynamics of Matter, 22761 Hamburg, Germany

[1] J. Waluk, *Chem. Rev.* **117**, 2447 (2017)

[2] Y. Litman, J. O. Richardson, T. Kumagai, and M. Rossi, J. Am. Chem. Soc. **141**, 2526 (2019)

[3] Y. Litman, J. Behler, and M. Rossi, Faraday Discuss. **221**, 526 (2020)

[4] Y. Litman and M. Rossi, Phys. Rev. Lett. **125,** 216001 (2020)

###### Register hereLink opens in a new window

**May 7, 2021, 1.00 pm BST**

**Accurate computation of core- and valence-level excitations of large systems from ***GW*-based methods

*GW*-based methods

**by Dr. Dorothea Golze**

###### Department of Applied Physics, Aalto University, 02150 Espoo, Finland

*GW*has become the method of choice for the calculation of valence photoemission spectra [1]. To apply

*GW*also to deep core excitations as measured by X-ray photoelectron spectroscopy (XPS), we recently advanced the

*GW*methodology and our implementation by combining exact numeric algorithms in the real frequency domain [2] with partial self-consistency [3] and relativistic corrections [3,4]. We benchmarked our core-level

*GW*implementation for small molecules and combined our

*GW*methodology with machine learning (ML) models to address also larger system sizes. We developed a powerful XPS prediction tool for materials and molecules containing carbon, hydrogen and oxygen by combining density functional theory and

*GW*calculations with Kernel Ridge Regression ML models. Another strategy to reach larger system sizes is to reduce the scaling of the

*GW*algorithms. In this talk, our latest work [5] on the development of an accurate low-scaling

*GW*algorithm and its application to valence-level excitations of phosphorene nanosheets will be presented.

[1] D. Golze, M. Dvorak, P. Rinke,

*Front. Chem*, 7 (2019), 377

[2] D. Golze, J. Wilhelm, M. van Setten, P. Rinke,

*JCTC*, 14 (2018), 4856

[3] D. Golze, L. Keller, P. Rinke,

*JPCL,*11 (2020), 1840

[4] L. Keller, V. Blum, P. Rinke, D. Golze,

*JCP*, 153 (2020), 114110[5] J. Wilhelm, P. Seewald, D. Golze,

*JCTC*, 17 (2021), 1662

###### Register hereLink opens in a new window

**June 4th, 2021, 1.00 pm BST**

**Generating 3d molecular structures with deep neural networks**

**by Niklas Gebauer**

###### Machine Learning Group, Technical University of Berlin, 10587 Berlin, Germany

Deep learning has proven to yield fast and accurate predictions of quantum-chemical properties to accelerate the discovery of novel molecules and materials.

As an exhaustive exploration of the vast chemical space is still infeasible, we require generative models that guide our search towards systems with desired properties.

While graph-based models have previously been proposed, they are restricted by a lack of spatial information such that they are unable to recognize spatial isomerism and non-bonded interactions.

Generative models for 3-dimensional structures are an emerging field promising to overcome these restrictions.

In this talk we will introduce the challenges of 3d molecule generation and focus on the autoregressive deep neural network G-SchNet which generates molecular structures by placing one atom after another in 3d space.

The model is capable of generating novel molecules that capture the characteristics of the training data and exhibit desired properties.