Calculating the "fingerprints" of molecules with artificial intelligence

The graphical neural network GNN receives small molecules as input with the task of determining their spectral responses. By matching them with the known spectra, the GNN programme learns to calculate spectra reliably.

The graphical neural network GNN receives small molecules as input with the task of determining their spectral responses. By matching them with the known spectra, the GNN programme learns to calculate spectra reliably. © K. Singh, A. Bande/HZB

With conventional methods, it is extremely time-consuming to calculate the spectral fingerprint of larger molecules. But this is a prerequisite for correctly interpreting experimentally obtained data. Now, a team at HZB has achieved very good results in significantly less time using self-learning graphical neural networks.

"Macromolecules but also quantum dots, which often consist of thousands of atoms, can hardly be calculated in advance using conventional methods such as DFT," says PD Dr. Annika Bande at HZB. With her team she has now investigated how the computing time can be shortened by using methods from artificial intelligence.

The idea: a computer programme from the group of "graphical neural networks" or GNN receives small molecules as input with the task of determining their spectral responses. In the next step, the GNN programme compares the calculated spectra with the known target spectra (DFT or experimental) and corrects the calculation path accordingly. Round after round, the result becomes better. The GNN programme thus learns on its own how to calculate spectra reliably with the help of known spectra.

"We have trained five newer GNNs and found that enormous improvements can be achieved with one of them, the SchNet model: The accuracy increases by 20% and this is done in a fraction of the computation time," says first author Kanishka Singh. Singh participates in the HEIBRiDS graduate school and is supervised by two experts from different backgrounds: computer science expert Prof. Ulf Leser from Humboldt University Berlin and theoretical chemist Annika Bande.

"Recently developed GNN frameworks could do even better," she says. "And the demand is very high. We therefore want to strengthen this line of research and are planning to create a new postdoctoral position for it from summer onwards as part of the Helmholtz project "eXplainable Artificial Intelligence for X-ray Absorption Spectroscopy"."

 

Annotation:

The work was carried out within the framework of the HEIBRiDS graduate school and is being supported by the Helmholtz project "eXplainable Artificial Intelligence for X-ray Absorption Spectroscopy" (XAI-4-XAS).

The core of the project is to extend GNN, as used at HZB, to very large molecules in combination with the probabilistic analysis of molecular motifs developed at HEREON. It is used to capture only the relevant part of the configuration phase space of the molecules, which is necessary for the accurate prediction of X-ray spectra. The results of the ML predictions allow a rigorous interpretation of XAS experiments, so that characteristic parts of the spectrum of an extended material can be assigned 1:1 to its specific structural subgroups.

 

arö

You might also be interested in

  • Quantum algorithms save time in the calculation of electron dynamics
    Science Highlight
    22.11.2022
    Quantum algorithms save time in the calculation of electron dynamics
    Quantum computers promise significantly shorter computing times for complex problems. But there are still only a few quantum computers worldwide with a limited number of so-called qubits. However, quantum computer algorithms can already run on conventional servers that simulate a quantum computer. A team at HZB has succeeded to calculate the electron orbitals and their dynamic development on the example of a small molecule after a laser pulse excitation. In principle, the method is also suitable for investigating larger molecules that cannot be calculated using conventional methods.
  • How photoelectrodes change in contact with water
    Science Highlight
    17.11.2022
    How photoelectrodes change in contact with water
    Photoelectrodes based on BiVO4 are considered top candidates for solar hydrogen production. But what exactly happens when they come into contact with water molecules? A study in the Journal of the American Chemical Society has now partially answered this crucial question:  Excess electrons from dopants or defects aid the dissociation of water which in turn stabilizes so-called polarons at the surface. This is shown by data from experiments conducted at the Advanced Light Source at Lawrence Berkeley National Laboratory. These insights might foster a knowledge-based design of better photoanodes for green hydrogen production.
  • BESSY II: Influence of protons on water molecules
    Science Highlight
    10.11.2022
    BESSY II: Influence of protons on water molecules
    How hydrogen ions or protons interact with their aqueous environment has great practical relevance, whether in fuel cell technology or in the life sciences. Now, a large international consortium at the X-ray source BESSY II has investigated this question experimentally in detail and discovered new phenomena. For example, the presence of a proton changes the electronic structure of the three innermost water molecules, but also has an effect via a long-range field on a hydrate shell of five other water molecules.