Machine Learning for Molecular Science

Jacob Mathias Schreiner: Machine Learning for Molecular Science

All the matter that we interact with in everyday life is made of fundamental building blocks called atoms. Atoms can configure in many ways into salts, molecules, crystals etc.. Chemistry is the study of how atoms reconfigure and rearrange into the matter we know.

The universe seeks to minimise energy and maximise entropy. These two factors are what drives chemistry at the microscopic level. Essentially, a chemical reaction transforming a set of reactants R into a set of products P happens because the configuration P has either lower energy or higher entropy. The trade-off between energy and entropy is governed by the temperature of the surrounding system. In this project we will mainly study the energy differences in chemical reactions.

Atoms have to physically move for chemical reactions to happen. This causes charged particles to come close even though they would naturally repel each other. Moving atoms in this way represents an energetic barrier and even if the end result of a chemical reaction might have a lower intrinsic energy than the initial state, the reaction might not happen if there is not enough energy available in the system to overcome this barrier. This high energy state in the middle of a reaction is called the transition state of the reaction.

A reaction path can be described in a 3N dimensional space - x, y and z coordinates for N atoms in a reaction. The path will start at a the initial configuration and end in the final configuration. There can be many paths in chemical space connecting these two states and they all have different energy barriers.  While some paths in this space would be impossible – if for example two or more atoms end up in the same region, the energy required would be enormous – others would have relatively low energy barriers. It just happens, that the rate of a certain reaction is exponentially proportional to the minimal energy barrier of all reaction paths.

Finding these transition states and thus calculating the rate of a reaction usually depends on human chemical intuition but here we want to make an automated framework that using generative modeling can predict reaction paths and transition states automatically.

PhD project

By: Jacob Mathias Schreiner

SectionCognitive Systems

Principal supervisor: Ole Winther

Co-supervisor: Tejs Vegge (DTU Energy)

Project titleMachine Learning for Molecular Science

Term: 01/09/2020 → 31/08/2023


Ole Winther
DTU Compute
+45 45 25 38 95


Tejs Vegge
Professor, Head of Section
DTU Energy
+45 45 25 82 01