Researchers from Skoltech and their colleagues from Germany and the United States studied the properties and behavior of a palladium-copper alloy under changing temperatures and hydrogen concentrations. The results of the research can be used for developing catalysts.
Transition metal-alloy materials can have catalytic properties and are therefore widely used in chemical reactions such as the carbon dioxide (CO2) hydrogenation and production of methanol from carbon dioxide.
Skoltech Center for Energy Science and Technology (CEST) researchers Zhong-Kang Han, Debalaya Sarker and Sergey Levchenko, together with colleagues from universities in Germany and the United States, modeled the properties of a palladium-copper alloy using machine learning to predict the distribution of palladium atoms on a copper surface depending on hydrogen partial pressure and temperature.
“Catalytically active sites on the surface are created only by palladium atoms, so it is important to know how many of these atoms will be on the surface at the appropriate temperatures and partial pressures of hydrogen,” Sergey Levchenko explains.
According to him, huge computational resources are required to evaluate the energies of many atomic configurations of palladium within the copper lattice in the presence of adsorbed hydrogen. Therefore, for this study, scientists decided to use a surrogate cluster expansion model that is more convenient to work with. “This model allows to evaluate the energy of millions of configurations in a matter of seconds”.
“Our system is much more complex than those that are usually studied using the method of cluster expansion. In this case, we investigated the alloy surface, where adsorbates from the gas phase affect the stability of various atomic configurations. That is why we used the machine learning method based on compressed sensing (it is widely used for image compression) and developed a highly accurate predictive surrogate model”, - Levchenko says.
The study was published in the Journal of Applied Physics
Ashmanov Neural Networks