MS project: Predicting the color of molecules

The color of a molecule (more precisely, its absorption wavelength) is determined by its chemical structure. However, the relationship between chemical structure and color remains elusive. Being able to establish this relationship—predicting the color from the chemical structure—would be valuable information to gain chemical insight.

To establish this structure-property relationship, considering basic aspects of molecular structures is useful: for instance, if a molecule contains a path with alternating single and double edges, “conjugated double bonds,” it tends to absorb light at longer wavelengths. Here, we aim to bring about further insight using a data-driven approach.

In this project, you will investigate to what extent we can use machine learning to predict the color of a molecule purely from its chemical structure. The use of adequate structural representations for molecular systems will be key in order to faithfully encode complex atom-in-molecule environments, while accounting for translational, rotational, and permutation invariance. You will train a machine learning model from a database of experimentally determined absorption and fluorescence wavelengths. The project will offer you a chance to work on state-of-the-art methods for machine learning applied to molecular systems.

Contact: Dr. Tristan Bereau ( Prof. Sander Woutersen (

MS project: Machine learning for polarizable force fields of organic molecules

The modeling of real materials, such as organic light emitting diodes and solar cells is inherently multiscale. For example, understanding interconversion of electrons, holes, and excitons requires a high resolution model, incorporating both the relevant physics and chemically-specific parameters. Polarizable force fields are ideally suited for this purpose. Unfortunately, their parametrization requires extensive manual tuning, at odds with a screening strategy that is desirable to design compounds with optimized properties.

Machine learning of multipole electrostatic parameters

The student will be involved in further developing efforts at the use of kernel-based machine learning (ML) methodologies to automate and ease parametrization for advanced force fields of organic molecules. He/she will be responsible for the use of covariant kernels for the learning of multipole electrostatic parameters across the chemical space of organic molecules. Previous work in this direction can be found in [1], [2], and [3]

[1] [2] [3]

We will extend this approach to the learning of multipole parameters for not only neutral organic molecules, but also anionic and cationic compounds. The student is expected to gain experience in building databases of electronic properties by means of high-throughput quantum-chemical calculations and training covariant kernels for physical properties.

Contact: Dr. Tristan Bereau ( Dr. Denis Andrienko (