direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

TU Berlin

Inhalt des Dokuments

Research Fields

Our group focuses on the development of Machine Learning (ML) algorithms for various problems in the natural sciences, ranging from quantum mechanical, atomistic systems to large scale industrial plant dynamics. We build domain specific modelswhich incorporate fundamental physical constraints and prior-knowledge to tackle various sorts of problems.

Prediction of chemical properties

Deep learning algorithms are powerful tools to model the complex relationships between the structure and properties of molecules and materials. The resulting models can be employed in a variety of tasks, ranging from the prediction of properties directly founded in quantum mechanics (e.g. molecular spectra) to more abstract concepts (e.g. toxicity). As such, they can aid in the discovery of new compounds as well as the rationalization of already established molecules and materials.

Inverse design – generation and optimization with respect to target properties

3-dimensional distribution over candidate atom positions from the G-SchNet model.

The space of all possible compounds is so vast that an exhaustive search for certain molecules or materials with desired properties is impossible, even if extremely fast evaluation of atomic structures was available at large-scale. Thus, inverse design aims at guiding the exploration of chemical compound space. Starting from a target property, a small set of new candidate structures shall be obtained in an informed way without needing to skim through millions of randomly assembled examples. We develop different ML approaches that allow inverse design either by optimizing a given structure or by generating new ones from scratch given some training examples.

Modelling and optimization of operational processes in industrial plants

Accurately modelling the dynamics of an industrual process can lead to a significantly more cost-efficient operation of the plant.
Instead of explicitly modelling the dynamics of the plant using a mechanistic model, which often proves difficult and costly, we develop recurrent neural networks (RNN) that are capable of efficiently learning a model of the process based on historic data. The RNNs allow us to forecast the plant operation for weeks and even months in advance, generate new synthetic sequences of the process based on the learned model, and can be extended to include domain knowledge about the specific process that can enable to networks to learn accurate models with less data.

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe