Theory – Computational Modeling & Materials Simulations (Pillar C)

Unit Heads

Key Researchers

Challenges

Modeling within the COE MECS needs to address complex processes, including catalytic
reactions, electronic excitations, and charge transport. Despite decades of research,
modeling these processes remains challenging due to trade-offs between cost and
accuracy. To solve and understand processes involving bond breaking, making, and
electronic excitations, it is in principle necessary to solve the time-dependent many-body
Schrödinger equation using correlated wavefunction methods. A computationally less
expensive method is density functional theory (DFT). However, DFT involves uncontrolled
approximations, which can for instance lead to the inaccurate description of excited states
or electronic states during a catalytic reaction involving strong multi-configurational
character.
Additionally, catalytic processes triggered by photons or electrons are often complex:
pristine surfaces are less reactive than step edges and corners, and in addition, dynamical
processes can take place on the surfaces involving complex reaction fronts that move
back and forth. This problem is well known under the name “size and complexity gap in
catalysis”. Addressing these challenges requires method development at multiple levels,
starting from the many-body Schrödinger equation for small systems, DFT at an
intermediate level, and bridging the scales to macroscopic observables using multi-scale
methods and machine learning (ML) techniques (Figure 1).

Figure 1: Towards realistic modeling of electro- and photo-catalytic systems, using a
combination of different computational approaches and developments. Existing methods
for treating excited states and periodic systems must be combined. Furthermore, machine
-learned force fields will be extended in order to treat species in different redox states,
electronically excited states as well as external potentials. Automatic structure search is
required to deal with increasingly complex systems. Finite temperature simulation will be
performed routinely to bridge the complexity gap.

Approach and Key topics

Ground and excited states in density functional theory

Modeling catalysis traditionally relies mainly on DFT and its variants for excited states instead of correlated wavefunction techniques due to their high computational demand. The Vienna ab initio simulation package (VASP) is pivotal for these calculations. Efforts will focus on calculating interatomic forces and non-adiabatic couplings in the excited state using DFT. The calculations will be constantly evaluated against more accurate methods, including RPA
and the GW approach, as well as multi-configurational methods. Specifically, we plan to
explore time-dependent DFT as well as restricted open-shell methods to follow the time-
evolution of periodic systems in the excited state.

Real-Time Dynamics in Many-Body theory

Dynamic simulations for photo-catalytic processes in real-time are crucial, requiring an
explicit solution of the time-dependent many-body Schrödinger equation. Techniques like
trajectory surface hopping (SHARC package) will be used, with plans to extend these non-
adiabatic molecular dynamics simulations from molecules to materials. The work will rely
on a combination of density functional theory (first key topic) and correlated wavefunction
techniques, where computationally feasible. The project aims to extend simulation time
scales to the nanosecond regime leveraging ML techniques developed in other sub-
projects.

First Principles Description of Complex Structures

Modeling catalytic processes requires treating bond-breaking and charge transfer
processes in complex structures (corners, edges, reaction fronts). To deal with this
complexity, embedding methods will be used to combine different levels of theory in a
single framework. Also, solvation models to model electro-catalysis at solid-liquid
interfaces will be extensively used, and interfaces to the SHARC package will be
developed.

Scale Bridging Methods and Machine-learned (ML) force fields

Machine learning techniques, in particular, ML force fields are an essential tool for bridging
the size and complexity gap in modeling interactions of molecules with complex substrates
or solvents. Approximate DFT calculations, on which ML force fields are often based, may
be inaccurate for predicting barriers, requiring correction methods like learning the
differences between different levels of theory or learning energy surfaces from correlated
wavefunction methods (transfer-learning).Structural complexity in energy materials demands highly transferable ML force fields and
methods to ensure correct asymptotic physical behavior. Combining these force fields with
automatic structure explorations and statistical methods to evaluate Gibbs free energies
and barriers is essential. Also, we plan to develop MLFFs that account for external
potentials and different charge states using highly accurate and transferable message-
passing neural networks. Ultimately, we plan to scale up the system size and simulation
time to reach realistic relevant time and length scales for catalytic systems.

Machine Learning Electronic Properties

The goal of this subproject is to describe the electronic structure of large heterostructures
without explicitly solving the DFT equation. This involves embedding ab-initio methods in
large-scale empirical approaches, and developing ML techniques to model the electronic
structure, transport and charge hopping in catalytic devices. This is among the most
challenging long-term goals of pillar C and will leverage expertise acquired in the previous
subprojects.