Home | People | Research | Courses | Publications | Address | News
Jobs | Master projects | Bachelor projects | Safety

Institute of Solid State Physics

      > > Deutsch     

Structure Search At Interfaces / Machine Learning      >> more >>

Description: We are currently developing our own algorithm to determine the geometric structure at inorganic/organic interfaces. This program employs machine learning techniques to establish the potential energy surface of organic/inorganic interfaces based on only a small number of density functional theory calculations. We are looking for highly motivated Bachelor students that are interested in computational physics, programming or method development and that would like to tackle projects that are directly relevant for our current research efforts. No specific skills or knowledge is required, but familiarity with Linux and basic programming experience (ideally Python) is recommended. Available topics include:

Explore, Expand, Exploit, Exterminate (Optimize Search Strategy)
Our algorithm explores the potential energy surface as evenly as possible. The data generated is then exploited (via machine learning) to sample and refine the predicted low-energy configurations. Both parts (exploration and exploitation) are computationally very expensive. The performance of the algorithm thus critically depends on finding a good ratio between those two, as well as using a good exploration strategy. Within this topic, we will critically assess exploration vs. exploitation for a variety of test systems, and determine a strategy to choose an optimal training set for a given budget of CPU time.

High-performance polymorphs (Machine Learning for various properties)
The same material can exhibit huge differences in its properties depending on the polymorph it assumes. To gauge the potential of a material for practical applications, it is often important to find what the “best” polymorph for a given property would be, even if it is not the energetically most favorable structure. The task in this project is to modify our algorithm such that it optimizes properties other than the energy and predicts, e.g., polymorphs with high interface dipoles or large charge-transport mobilities.

Molecular Lego
Finding out how molecules adsorb on a surface is a challenging task on its own. Research groups all around the world are developing methods to exploit the potential energy surface with advanced methods such as genetic algorithms, basin hopping and Gaussian process regression. We take this task a step further and project the energies onto meaningful building blocks of the molecule. This allows us to identify the main interacting parts of the molecules and in the future exchange building blocks to favorize specific adsorption positions. We search for bachelor students who want to improve the algorithm or apply it to a new set of molecules. 

Transition Rates and Barriers (Metastable Phases)
Some of the most interesting materials are not thermodynamically stable, but only kinetically trapped by (more or less) large reaction barriers. Developing new materials therefore requires not only finding structures with outstanding properties, but also knowledge about their expected lifetime. A rough estimate of the barrier height can be obtained by calculating the vibrational properties of different structures and determining their intersection. The task of this bachelor thesis is to assess the quality of this approximation for a selected example

Hessian Learning and Geometry Optimization
One of the key factors determining the efficiency of geometry optimization at the quantum-mechanical level is the initial guess of the Hesse matrix, i.e. the second derivative of energy with respect to atomic displacements. Most contemporary algorithms are designed for molecules or solid crystals in mind, but yield poor results for interfaces. The task of the present topic is to use machine-learning approaches to create an improved Hesse-Matrix guess for specific polymorphs based on pre-calculated Hesse-Matrices of other structures.

Teaching Atom Species to Machine Learning
Let us assume we have two molecules, i.e. a dimer, and wish to know their interaction energy: Of course this energy depends on their relative positions and orientations. We can calculate the energy for a limited number of positions and orientations, but attaining the full picture requires the use of machine leaning. Machine learning of molecular interactions relies on descriptors that depend on the relative positions and orientations and allow us to compare different dimers. We have developed such a descriptor based on radial distance functions. However, as of yet it does not account for different atoms species. This is where you come in! Help us to improve our machine learning model by implementing atom species dependence.

Oliver Hofmann
email: o.hofmann@tugraz.at
Tel: 0316 873 8964
Or talk to the students in office PH 02 152 (2nd floor, right by the stairs)