HomeMachine LearningMachine Learning NewsScientist are using Machine Learning toadvance simulation of metal-organic frameworks

Scientist are using Machine Learning toadvance simulation of metal-organic frameworks

Metal-organic frameworks, or MOFs, exhibit remarkable capabilities such as heat conduction, water sequestration, gas storage, CO2 storage, and hydrogen storage because of their special structure in the form of microporous crystals, which are microscopic in size but have a huge surface area. They are therefore quite intriguing for both practical use and research. However, because MOFs are extremely complex systems, successful simulation has thus far required a significant investment of time and computational power.

The creation and implementation of innovative MOFs has been dramatically accelerated by a team lead by Egbert Zojer from the Institute of Solid State Physics at Graz University of Technology (TU Graz), who has now used machine learning to significantly improve these simulations. The technique used by the researchers has been published in npj Computational Materials.

It was previously impossible to simulate with the precision of quantum mechanical techniques.

It is required for simulating massive supercells in order to simulate specific MOF features. This pertains, for instance, to the computation of heat conduction in MOFs, which is extremely important for practically all applications “Egbert Zojer outlines the difficulty that needed to be overcome.

Tens of thousands or even hundreds of thousands of atoms are frequently found in the simulated supercells. Then, five to ten million times must be spent solving the equations of motion for these massive systems. This is considerably beyond what can currently be computed with robust quantum mechanical techniques.

For such computations, transferable force fields that were frequently parametrized based on tests were thus frequently employed up until this point. Nevertheless, the outcomes acquired using these force fields proved to be mostly insufficiently trustworthy.

The utilization of machine-learned potentials has now profoundly transformed this. By using a newly devised interplay of existing algorithms, including methods established at the University of Vienna, they are applied to quantum mechanical simulations. Quantum mechanical simulations need to be performed for a relatively limited number of much smaller structures in order to do the required material-specific machine learning of the potentials.

Consequently, this leads to orders of magnitude faster calculation times and allows modern supercomputers to model the forces in the massive supercells millions of times. The key benefit here is that, in comparison to performing the simulations using quantum mechanical techniques, there is no appreciable loss of accuracy.

More effective way to look for the properties you want

This means that, in the case of the heat conduction of MOFs, the recently created modeling approach will enable the simulation of pertinent material properties even prior to the synthesis of MOFs, enabling researchers to consistently create tailored structures on a computer.

This is a significant advancement in the study of complex materials and will, for example, enable researchers to optimize the interaction between metal oxide nodes and semiconducting organic linkers in heat transport. Overcoming difficult obstacles will also be simpler when the new simulation technique is used. For instance, depending on their intended use, MOFs may need to have excellent or poor thermal conductivity.

For example, a hydrogen storage system needs to have good heat dissipation capabilities, while thermoelectric applications need to combine low heat dissipation with good electrical conduction.

The new machine-learned potentials are perfect for computing MOFs’ various dynamic and structural properties, in addition to simulating heat conductivity. These include phonons, vibrational spectra, elastic constants, and crystalline structures, all of which are critical to the charge transport characteristics and thermal stability of MOFs.

We now possess instruments that we are confident are highly effective in giving us accurate quantitative data. Because of this, we may systematically alter the MOFs’ structures in the simulations and yet be confident in the accuracy of the predicted properties. According to Egbert Zojer, this will enable us to determine whether atomistic structure modifications produce the intended effects based on causality. Despite the novel simulation strategy’s recent publication, research groups in Munich and Bayreuth have already adopted it.

Source link

Most Popular