Accelerating Non-Empirical Structure Determination of Ziegler-Natta Catalysts with a High-Dimensional Neural Network Potential

The determination of catalyst nanostructures with first-principlesaccuracy using genetic algorithms (GA) is very demanding due to thecubic scaling of the computational cost of density functional theory(DFT) calculations. Here, we demonstrate, for the case of Ziegler-NattaMgCl(2)/TiCl4 nanoplates, how this structure determinationcan be accelerated by employing a high-dimensional neural networkpotential (HDNNP) of essentially DFT accuracy. First, when buildingHDNNPs for MgCl2/TiCl4 clusters with computationallytractable sizes, we found that the structural diversity in the trainingset is crucial for obtaining HDNNPs reliably describing the largevariety of structures generated by GA. The resulting HDNNPs dramaticallyaccelerated the structure determination while yielding results consistentwith DFT. Subsequently, we developed a multistep adaptive procedureto construct a HDNNP for MgCl2/TiCl4 clustersconsistent in size and TiCl4 coverage with experimentswhere prior DFT results were scarcely collected. The structure determinationand analyses underline the importance of system size and compositionin order to predict some experimentally known facts such as the surfacemorphology and population of isospecific sites.