direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

TU Berlin

Inhalt des Dokuments

Prediction of chemical properties

Currently, chemical innovation is guided by painstaking experimentation and the expert knowledge of chemists. However, the emergence of powerful deep learning architectures and extensive databases of experimental and computational information on chemical compounds has opened alternative routes towards the rational design of functional compounds. Machine learning models can extract and distill the knowledge encoded in these databases, learning the complex relationships between chemical structure and properties. These models can then be used guide the search for new materials and properties or aid in the characterization and rationalization of known chemicals. In BASLEARN, we aim to create deep learning architectures that are ideally suited for these tasks, incorporating fundamental quantum-chemical concepts on an atomistic level, enabling us to accurately and efficiently predict the properties of new chemical compounds.

One such architecture is SchNet, a neural network specifically designed specifically for atomistic systems. SchNet models interactions between atoms using continuous-filter convolutions, producing atom-wise descriptors which encode chemical and structural information. Due to its atom-centric design, the architecture is capable of accurately predicting a wide range of properties across chemical compound space for molecules and materials, as well as modeling potential-energy surfaces and energy-conserving force fields for molecular dynamics simulations.

An extension of SchNet, SchNOrb (SchNet for Orbitals) was designed for the prediction of the quantum mechanical wavefunction, from which a wealth of other properties can be derived directly. This approach retains full access to the electronic structure via the wavefunction at force-field-like efficiency and captures quantum mechanics in an analytically differentiable representation. This architecture opens promising avenues to perform inverse design of molecular structures for target electronic property optimization.

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe